@bendyline/gezel-inference 1.0.5 → 1.0.7

package/dist/catalog/image-loras/sd/sdxl-coloring-book/manifest.json

@@ -0,0 +1,30 @@
+{
+  "schemaVersion": 1,
+  "kind": "image-lora",
+  "id": "sdxl-coloring-book",
+  "name": "Coloring book",
+  "description": "Draws clean black line art to color in.",
+  "tags": [
+    "line-art",
+    "coloring-book"
+  ],
+  "maintainer": {
+    "name": "artificialguybr",
+    "url": "https://huggingface.co/artificialguybr/ColoringBookRedmond-V2"
+  },
+  "license": "CreativeML Open RAIL-M",
+  "licenseClass": "custom-restricted",
+  "licenseShortName": "OpenRAIL-M",
+  "licenseUrl": "https://huggingface.co/spaces/CompVis/stable-diffusion-license",
+  "yankedVersions": [],
+  "baseModels": [
+    "sdxl-lightning-4step"
+  ],
+  "purpose": "style",
+  "strength": 1.4,
+  "prompt": {
+    "prefix": "ColoringBookAF, Coloring Book, black and white line art,"
+  },
+  "upstream": "https://huggingface.co/artificialguybr/ColoringBookRedmond-V2",
+  "commercialUse": true
+}

package/dist/catalog/image-loras/sd/sdxl-coloring-book/versions/1.0.0/manifest.json

@@ -0,0 +1,15 @@
+{
+  "schemaVersion": 1,
+  "version": "1.0.0",
+  "releasedAt": "2026-10-31T00:00:00Z",
+  "downloadUrl": "https://huggingface.co/artificialguybr/ColoringBookRedmond-V2/resolve/main/ColoringBookRedmond-ColoringBook-ColoringBookAF.safetensors",
+  "sha256": "94eb17068a45df06a9e6ab0bcbbd5c381e03f09320da46839c1ed0cee733af31",
+  "approxSizeBytes": 170540036,
+  "tensors": {
+    "floatElements": {
+      "BF16": 85115602
+    },
+    "otherBytes": 0
+  },
+  "notes": "Benched 2026-10-31 on an RTX Spark N1X (Windows ARM64, CUDA, sd-cpp master-975) on SDXL Lightning 4-step: clean black-and-white line art on all four smoke prompts at 1.4 with \"black and white line art,\" added, even worded as photos; at 1.0 some color and shading leak through. Lettering breaks."
+}

package/dist/catalog/image-loras/sd/sdxl-papercut/manifest.json

@@ -0,0 +1,29 @@
+{
+  "schemaVersion": 1,
+  "kind": "image-lora",
+  "id": "sdxl-papercut",
+  "name": "Paper cut",
+  "description": "Builds the picture from layers of cut paper.",
+  "tags": [
+    "paper-cut"
+  ],
+  "maintainer": {
+    "name": "TheLastBen",
+    "url": "https://huggingface.co/TheLastBen/Papercut_SDXL"
+  },
+  "license": "CreativeML Open RAIL-M",
+  "licenseClass": "custom-restricted",
+  "licenseShortName": "OpenRAIL-M",
+  "licenseUrl": "https://huggingface.co/spaces/CompVis/stable-diffusion-license",
+  "yankedVersions": [],
+  "baseModels": [
+    "sdxl-lightning-4step"
+  ],
+  "purpose": "style",
+  "strength": 1,
+  "prompt": {
+    "prefix": "papercut,"
+  },
+  "upstream": "https://huggingface.co/TheLastBen/Papercut_SDXL",
+  "commercialUse": true
+}

package/dist/catalog/image-loras/sd/sdxl-papercut/versions/1.0.0/manifest.json

@@ -0,0 +1,15 @@
+{
+  "schemaVersion": 1,
+  "version": "1.0.0",
+  "releasedAt": "2026-10-31T00:00:00Z",
+  "downloadUrl": "https://huggingface.co/TheLastBen/Papercut_SDXL/resolve/main/papercut.safetensors",
+  "sha256": "98f1729d858e7bf719283797e45661337ce915941b83260026e157d94963c402",
+  "approxSizeBytes": 340786114,
+  "tensors": {
+    "floatElements": {
+      "F16": 170230482
+    },
+    "otherBytes": 0
+  },
+  "notes": "Benched 2026-10-31 on an RTX Spark N1X (Windows ARM64, CUDA, sd-cpp master-975) on SDXL Lightning 4-step: a strong layered paper look on all four smoke prompts, adding paper frames and breaking lettering; 0.8 looked the same."
+}

package/dist/catalog/image-loras/sd/sdxl-pixel-art/manifest.json

@@ -0,0 +1,29 @@
+{
+  "schemaVersion": 1,
+  "kind": "image-lora",
+  "id": "sdxl-pixel-art",
+  "name": "Pixel art",
+  "description": "Draws in chunky pixel art.",
+  "tags": [
+    "pixel-art"
+  ],
+  "maintainer": {
+    "name": "nerijs",
+    "url": "https://huggingface.co/nerijs/pixel-art-xl"
+  },
+  "license": "CreativeML Open RAIL-M",
+  "licenseClass": "custom-restricted",
+  "licenseShortName": "OpenRAIL-M",
+  "licenseUrl": "https://huggingface.co/spaces/CompVis/stable-diffusion-license",
+  "yankedVersions": [],
+  "baseModels": [
+    "sdxl-lightning-4step"
+  ],
+  "purpose": "style",
+  "strength": 1,
+  "prompt": {
+    "prefix": "pixel art,"
+  },
+  "upstream": "https://huggingface.co/nerijs/pixel-art-xl",
+  "commercialUse": true
+}

package/dist/catalog/image-loras/sd/sdxl-pixel-art/versions/1.0.0/manifest.json

@@ -0,0 +1,15 @@
+{
+  "schemaVersion": 1,
+  "version": "1.0.0",
+  "releasedAt": "2026-10-31T00:00:00Z",
+  "downloadUrl": "https://huggingface.co/nerijs/pixel-art-xl/resolve/main/pixel-art-xl.safetensors",
+  "sha256": "4234637cb80c998f41e348e6a6cb6bc20d8d038b2b0f256b6129b3b5e353eef7",
+  "approxSizeBytes": 170543052,
+  "tensors": {
+    "floatElements": {
+      "BF16": 85115602
+    },
+    "otherBytes": 0
+  },
+  "notes": "Benched 2026-10-31 on an RTX Spark N1X (Windows ARM64, CUDA, sd-cpp master-975) on SDXL Lightning 4-step: convincing pixel art on the apple and portrait, a tiled pattern on the cubes and broken lettering on the sign; about 1.6 s more an image."
+}

package/dist/catalog/image-models/fl/flux-2-klein-4b-q4/manifest.json

@@ -38,6 +38,9 @@
   "promptStyle": "sentences",
   "rendersText": true,
   "supportsImg2Img": false,
+  "referenceImages": {
+    "max": 3
+  },
   "hardwareTier": "high",
   "minRamGB": 16,
   "commercialUse": true

package/dist/catalog/image-models/index.json

@@ -1 +1 @@
-{"schemaVersion":1,"kind":"image-model","count":13,"entries":[{"manifest":{"schemaVersion":1,"kind":"image-model","id":"animagine-xl-4","name":"Animagine XL 4.0","description":"Cagliostro Lab's flagship open anime/illustration model. SDXL-compatible so all SDXL tooling and LoRAs work. Booru-style tag prompts produce the best results. The strongest open anime model available.","tags":["sdxl","anime","illustration","artistic"],"maintainer":{"name":"Cagliostro Lab","url":"https://huggingface.co/cagliostrolab/animagine-xl-4.0"},"license":"Fair AI Public License 1.0-SD","licenseClass":"custom-restricted","licenseShortName":"Fair AI 1.0-SD","licenseUrl":"https://huggingface.co/cagliostrolab/animagine-xl-4.0","version":"4.0.1","releasedAt":"2026-10-11T00:00:00Z","downloadUrl":"https://huggingface.co/cagliostrolab/animagine-xl-4.0/resolve/main/animagine-xl-4.0.safetensors","sha256":"1d5b43ff75b6ab598502d4c779d2fbfa3dceca51c60c3b609640a60772333916","approxSizeBytes":6938434056,"quantization":"FP16","recommendedSteps":28,"widthRange":{"min":512,"max":1536},"heightRange":{"min":512,"max":1536},"upstream":"https://huggingface.co/cagliostrolab/animagine-xl-4.0","category":"artistic","weightsKind":"checkpoint","supportsImg2Img":true,"auxiliaryFiles":[],"hardwareTier":"mid","minRamGB":10,"commercialUse":true,"availableVersions":["4.0.1","4.0.0"]}},{"manifest":{"schemaVersion":1,"kind":"image-model","id":"flux-2-klein-4b-q4","name":"FLUX.2 Klein 4B (Q4)","description":"Black Forest Labs' FLUX.2-klein-4B — the successor to FLUX.1-schnell. A 4-step, CFG-free, 4B-parameter rectified flow transformer with state-of-the-art quality on consumer hardware. This package needs about 8 GB of memory and renders a 1024px image in about 7 seconds on a DGX Spark. Drops the FLUX.1 T5+CLIP encoder pair for a single Qwen3-4B LLM encoder. leejet's Q4_0 quantization of the diffusion model paired with Qwen3-4B-Q4_K_M and the FLUX.2-small-decoder VAE. Apache 2.0 throughout — fully commercial-friendly.","tags":["flux","flux2","flagship","fast","4-step","apache","llm-encoder"],"maintainer":{"name":"Black Forest Labs (quantized by leejet)","url":"https://huggingface.co/leejet/FLUX.2-klein-4B-GGUF"},"license":"Apache 2.0","licenseClass":"open","licenseShortName":"Apache 2.0","licenseUrl":"https://huggingface.co/black-forest-labs/FLUX.2-klein-4B/blob/main/LICENSE.md","recoScore":20,"version":"1.0.1","releasedAt":"2026-10-11T00:00:00Z","downloadUrl":"https://huggingface.co/leejet/FLUX.2-klein-4B-GGUF/resolve/main/flux-2-klein-4b-Q4_0.gguf","sha256":"d1023499ef3f2f82ff7c50e6778495195c1b6cc34835741778868428111f9ff4","approxSizeBytes":2460378560,"quantization":"Q4_0","recommendedSteps":4,"widthRange":{"min":512,"max":2048},"heightRange":{"min":512,"max":2048},"upstream":"https://huggingface.co/black-forest-labs/FLUX.2-klein-4B","category":"general","weightsKind":"diffusion-model","supportsImg2Img":false,"auxiliaryFiles":[{"role":"vae","downloadUrl":"https://huggingface.co/black-forest-labs/FLUX.2-small-decoder/resolve/main/full_encoder_small_decoder.safetensors","sha256":"ea4273f02d1fafbf8e1d1c2cf6018ed8748652eb0bf34f2dd91171f16f15ab62","approxSizeBytes":249519092,"tensors":{"floatElements":{"F32":62373347},"otherBytes":8}},{"role":"llm","downloadUrl":"https://huggingface.co/unsloth/Qwen3-4B-GGUF/resolve/main/Qwen3-4B-Q4_K_M.gguf","sha256":"f6f851777709861056efcdad3af01da38b31223a3ba26e61a4f8bf3a2195813a","approxSizeBytes":2497281312}],"hardwareTier":"high","minRamGB":16,"commercialUse":true,"availableVersions":["1.0.1","1.0.0"]}},{"manifest":{"schemaVersion":1,"kind":"image-model","id":"iris-3b-f32","name":"Iris 3B (F32)","description":"Generate photographs, illustrations, and artwork at about one megapixel. Iris paints pixels directly, with a separate Qwen text encoder and no VAE decoder. Approximately 14.5 GB of downloads. It renders at 50 steps, about 2.5 minutes per image on a DGX Spark.","tags":["iris","text-to-image","pixel-space","photography","illustration"],"maintainer":{"name":"Sperid Labs","url":"https://huggingface.co/speridlabs"},"license":"Apache-2.0","licenseClass":"open","licenseShortName":"Apache 2.0","licenseUrl":"https://www.apache.org/licenses/LICENSE-2.0","minGezelVersion":"1.26283","version":"1.0.2","releasedAt":"2026-10-19T00:00:00Z","downloadUrl":"https://huggingface.co/speridlabs/iris-3b/resolve/7445443349bc9abe3c96f01ff793e2098ca012b3/model.safetensors","sha256":"a90e1f4c197bac8491f82078def68884b6a3527ea7fc97a693e3e367372aa80f","approxSizeBytes":11950112680,"quantization":"F32","recommendedSteps":50,"widthRange":{"min":512,"max":1536},"heightRange":{"min":512,"max":1536},"upstream":"https://huggingface.co/speridlabs/iris-3b","category":"general","weightsKind":"diffusion-model","supportsImg2Img":false,"auxiliaryFiles":[{"role":"llm","downloadUrl":"https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct-GGUF/resolve/1cd86afb9a95c410a6038ab3b40d8b578c892266/Qwen3VL-4B-Instruct-Q4_K_M.gguf","sha256":"66358cb18bb6b3b1b6675aa412c7a88ef01d228f481184d13668e5201c730a0a","approxSizeBytes":2497281664}],"hardwareTier":"high","minRamGB":32,"commercialUse":true,"notes":"Requires a Gezel build with Iris defaults and stable-diffusion.cpp master-951-f89d9b1 or newer. Original F32 image weights are loaded in BF16; the Qwen3-VL-4B-Instruct Q4_K_M text encoder keeps its quantization. Use CFG 3, Euler, diffusion flash attention, and dimensions divisible by 16 at approximately one megapixel. The publisher recommends 100 steps; Gezel renders 50, which matched 100 in the Gezel image bench at half the time. A 20-step 1024x1024 render took 174 seconds on a 128 GB Apple M5 Max, with about 11.8 GB peak resident memory. On a DGX Spark (GB10, CUDA) a 1024px render took about 150 seconds at 50 steps and 300 at 100, with a peak of 13 to 15 GiB of memory, so 16 GB machines are tight. Captions are limited to 300 tokens including the assistant-turn suffix. No VAE or vision projector is needed. Text-to-image only: depth estimation and restoration/upscaling require separate checkpoints and are not included.","availableVersions":["1.0.2","1.0.1","1.0.0"]}},{"manifest":{"schemaVersion":1,"kind":"image-model","id":"krea-2-turbo-q4","name":"Krea 2 Turbo (Q4)","description":"Krea's flagship Krea 2 Turbo — a 12B diffusion transformer distilled for fast 8-step generation at 1K-2K resolution, paired with the Qwen-Image VAE and a Qwen3-VL text encoder. Abiray's Q4_K_M quantization is the consumer-hardware tier. Non-commercial (Krea 2 Community License).","tags":["krea","krea-2","flagship","8-step","distilled","non-commercial"],"maintainer":{"name":"Krea (quantized by Abiray)","url":"https://huggingface.co/Abiray/Krea-2-Turbo-GGUF"},"license":"Krea 2 Community License","licenseClass":"commercial-restricted","licenseShortName":"Krea 2 Community","licenseUrl":"https://huggingface.co/krea/Krea-2-Turbo/blob/main/LICENSE.pdf","version":"1.0.1","releasedAt":"2026-10-11T00:00:00Z","downloadUrl":"https://huggingface.co/Abiray/Krea-2-Turbo-GGUF/resolve/main/Krea-2-Turbo-Q4_K_M.gguf","sha256":"1e23e478144ce4c1adc6854efe2ff52930a836cf4e0008b6a0e41d235d9bdd07","approxSizeBytes":7486289232,"quantization":"Q4_K_M","recommendedSteps":8,"widthRange":{"min":512,"max":2048},"heightRange":{"min":512,"max":2048},"upstream":"https://huggingface.co/krea/Krea-2-Turbo","category":"general","weightsKind":"diffusion-model","supportsImg2Img":false,"auxiliaryFiles":[{"role":"vae","downloadUrl":"https://huggingface.co/Comfy-Org/Krea-2/resolve/main/vae/qwen_image_vae.safetensors","sha256":"a70580f0213e67967ee9c95f05bb400e8fb08307e017a924bf3441223e023d1f","approxSizeBytes":253806246,"tensors":{"floatElements":{"BF16":126892531},"otherBytes":0}},{"role":"llm","downloadUrl":"https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct-GGUF/resolve/main/Qwen3VL-4B-Instruct-Q4_K_M.gguf","sha256":"66358cb18bb6b3b1b6675aa412c7a88ef01d228f481184d13668e5201c730a0a","approxSizeBytes":2497281664}],"hardwareTier":"high","minRamGB":16,"commercialUse":false,"availableVersions":["1.0.1","1.0.0"]}},{"manifest":{"schemaVersion":1,"kind":"image-model","id":"krea-2-turbo-q6","name":"Krea 2 Turbo (Q6)","description":"Krea's flagship Krea 2 Turbo — a 12B diffusion transformer distilled for fast 8-step generation at 1K-2K resolution, paired with the Qwen-Image VAE and a Qwen3-VL text encoder. Abiray's Q6_K quantization trades extra VRAM for near-Q8 fidelity. Non-commercial (Krea 2 Community License).","tags":["krea","krea-2","flagship","8-step","distilled","non-commercial"],"maintainer":{"name":"Krea (quantized by Abiray)","url":"https://huggingface.co/Abiray/Krea-2-Turbo-GGUF"},"license":"Krea 2 Community License","licenseClass":"commercial-restricted","licenseShortName":"Krea 2 Community","licenseUrl":"https://huggingface.co/krea/Krea-2-Turbo/blob/main/LICENSE.pdf","version":"1.0.1","releasedAt":"2026-10-11T00:00:00Z","downloadUrl":"https://huggingface.co/Abiray/Krea-2-Turbo-GGUF/resolve/main/Krea-2-Turbo-Q6_K.gguf","sha256":"91103da0cf8e2b79f2108d4c6b885a62bacf7911c633373bbe0556b6da9a1f5c","approxSizeBytes":10583638350,"quantization":"Q6_K","recommendedSteps":8,"widthRange":{"min":512,"max":2048},"heightRange":{"min":512,"max":2048},"upstream":"https://huggingface.co/krea/Krea-2-Turbo","category":"general","weightsKind":"diffusion-model","supportsImg2Img":false,"auxiliaryFiles":[{"role":"vae","downloadUrl":"https://huggingface.co/Comfy-Org/Krea-2/resolve/main/vae/qwen_image_vae.safetensors","sha256":"a70580f0213e67967ee9c95f05bb400e8fb08307e017a924bf3441223e023d1f","approxSizeBytes":253806246,"tensors":{"floatElements":{"BF16":126892531},"otherBytes":0}},{"role":"llm","downloadUrl":"https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct-GGUF/resolve/main/Qwen3VL-4B-Instruct-Q4_K_M.gguf","sha256":"66358cb18bb6b3b1b6675aa412c7a88ef01d228f481184d13668e5201c730a0a","approxSizeBytes":2497281664}],"hardwareTier":"workstation","minRamGB":24,"commercialUse":false,"availableVersions":["1.0.1","1.0.0"]}},{"manifest":{"schemaVersion":1,"kind":"image-model","id":"krea-2-turbo-q8","name":"Krea 2 Turbo (Q8)","description":"Krea's flagship Krea 2 Turbo — a 12B diffusion transformer distilled for fast 8-step generation at 1K-2K resolution, paired with the Qwen-Image VAE and a Qwen3-VL text encoder. Abiray's Q8_0 quantization is near-lossless for workstation-class hardware. Non-commercial (Krea 2 Community License).","tags":["krea","krea-2","flagship","8-step","distilled","non-commercial"],"maintainer":{"name":"Krea (quantized by Abiray)","url":"https://huggingface.co/Abiray/Krea-2-Turbo-GGUF"},"license":"Krea 2 Community License","licenseClass":"commercial-restricted","licenseShortName":"Krea 2 Community","licenseUrl":"https://huggingface.co/krea/Krea-2-Turbo/blob/main/LICENSE.pdf","version":"1.0.1","releasedAt":"2026-10-11T00:00:00Z","downloadUrl":"https://huggingface.co/Abiray/Krea-2-Turbo-GGUF/resolve/main/Krea-2-Turbo-Q8_0.gguf","sha256":"8be11f7fa2b63cab28be213246001195ebe04f271a703b351461e0b775e3f775","approxSizeBytes":13705958734,"quantization":"Q8_0","recommendedSteps":8,"widthRange":{"min":512,"max":2048},"heightRange":{"min":512,"max":2048},"upstream":"https://huggingface.co/krea/Krea-2-Turbo","category":"general","weightsKind":"diffusion-model","supportsImg2Img":false,"auxiliaryFiles":[{"role":"vae","downloadUrl":"https://huggingface.co/Comfy-Org/Krea-2/resolve/main/vae/qwen_image_vae.safetensors","sha256":"a70580f0213e67967ee9c95f05bb400e8fb08307e017a924bf3441223e023d1f","approxSizeBytes":253806246,"tensors":{"floatElements":{"BF16":126892531},"otherBytes":0}},{"role":"llm","downloadUrl":"https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct-GGUF/resolve/main/Qwen3VL-4B-Instruct-Q4_K_M.gguf","sha256":"66358cb18bb6b3b1b6675aa412c7a88ef01d228f481184d13668e5201c730a0a","approxSizeBytes":2497281664}],"hardwareTier":"workstation","minRamGB":32,"commercialUse":false,"availableVersions":["1.0.1","1.0.0"]}},{"manifest":{"schemaVersion":1,"kind":"image-model","id":"ming-image-0.1-design-bf16","name":"Ming Image 0.1 Design (BF16)","description":"Generate posters, infographics, UI mockups, and text-rich visual designs locally. The 6B image transformer and separate Ling text encoder use BF16 weights. A large workstation model: approximately 49 GB of downloads; memory and speed depend on resolution and offloading.","tags":["ming","design","infographics","posters","bf16"],"maintainer":{"name":"inclusionAI (packaged by Comfy-Org)","url":"https://huggingface.co/Comfy-Org/Ming-Image"},"license":"MIT","licenseClass":"open","licenseShortName":"MIT","licenseUrl":"https://huggingface.co/inclusionAI/Ming-Image-0.1-Design/blob/main/LICENSE","minGezelVersion":"1.26283","version":"1.0.2","releasedAt":"2026-10-19T00:00:00Z","downloadUrl":"https://huggingface.co/Comfy-Org/Ming-Image/resolve/93804ff3747517a318969c5f61b25a57daefe09f/diffusion_models/ming_image_0.1_design_bf16.safetensors","sha256":"8781c6fc679f1812ad318cc7fd84c4d05a23b3571b9a7efeff73a92054ef4659","approxSizeBytes":12309858664,"quantization":"BF16","recommendedSteps":12,"widthRange":{"min":1024,"max":2048},"heightRange":{"min":1024,"max":2048},"upstream":"https://huggingface.co/inclusionAI/Ming-Image-0.1-Design","category":"general","weightsKind":"diffusion-model","supportsImg2Img":false,"auxiliaryFiles":[{"role":"vae","downloadUrl":"https://huggingface.co/Comfy-Org/Ming-Image/resolve/93804ff3747517a318969c5f61b25a57daefe09f/vae/ming_image_vae_bf16.safetensors","sha256":"7f5bed402dc8c77dc2e0ab1929a85d4df433b7cf7b599dfa8c353da98db0b90a","approxSizeBytes":253816696,"tensors":{"floatElements":{"BF16":126897716},"otherBytes":0}},{"role":"llm","downloadUrl":"https://huggingface.co/Comfy-Org/Ming-Image/resolve/93804ff3747517a318969c5f61b25a57daefe09f/text_encoders/ming_image_0.1_ling_mini_2.0_bf16.safetensors","sha256":"7eccad21934f2e6d9103b5c589a448082e89f7b5b73a6b9c6d0332be7de26c1e","approxSizeBytes":36677768245,"tensors":{"floatElements":{"BF16":18332699136,"F32":9728},"otherBytes":12210709}},{"role":"tokenizer","downloadUrl":"https://huggingface.co/inclusionAI/Ming-Image-0.1-Design/resolve/208087ada1486931692c1896f38d4cd16ff3df82/mllm/tokenizer.json","sha256":"e7ff01708d504f7bf4dbf7f5815adde57bab9a40e7f563ab6ad1acace4464917","approxSizeBytes":12210709}],"hardwareTier":"workstation","minRamGB":96,"commercialUse":true,"notes":"Requires a Gezel build with Ming support and stable-diffusion.cpp master-951-f89d9b1 or newer. Use 12 steps, CFG 1, Euler, and dimensions divisible by 16. Keep the Ling text encoder in BF16. Text-to-image only; the separate Design-Layer model is not included. BF16 avoids the missing INT8 matrix kernel observed on Metal with this engine pin. A 1024px design poster passed end-to-end testing on a 128 GB Apple M5 Max; memory guidance is conservative, not a measured minimum. Local generation took approximately 5–7 minutes with VAE tiling. On a DGX Spark (GB10, CUDA) a 1024px image took about 24 seconds, with a peak of about 46 GiB of memory. Transparent-background output is not validated on this Metal build: both a generic transparency request and the documented RGBA prefix produced opaque white backgrounds, while PNG bytes were preserved unchanged. Upstream transparency guidance: https://github.com/inclusionAI/Ming-Image#transparent-background-generation-tip.","availableVersions":["1.0.2","1.0.1","1.0.0"]}},{"manifest":{"schemaVersion":1,"kind":"image-model","id":"moss-image-sft-q4","name":"MOSS-Image SFT (Q4)","description":"OpenMOSS's MOSS-Image, fine-tuned to follow instructions. Strong at clear, accurate text inside images (signs, posters, whiteboards, labeled diagrams, Chinese as well as English), and it can place text and objects where a prompt or a JSON layout says. A 5B image transformer with Qwen3.5 9B as its text encoder and the FLUX.2 small-decoder VAE: leejet's Q4_K quantization of the transformer and a Q4_K_M encoder, about 9.3 GB of downloads. Careful rather than fast: 50 steps with guidance means about a hundred transformer passes per image. Apache 2.0 throughout, so fully commercial-friendly.","tags":["moss","text-rendering","layouts","apache","llm-encoder"],"maintainer":{"name":"OpenMOSS (quantized by leejet)","url":"https://huggingface.co/leejet/MOSS-Image-SFT-GGUF"},"license":"Apache 2.0","licenseClass":"open","licenseShortName":"Apache 2.0","licenseUrl":"https://huggingface.co/OpenMOSS-Team/MOSS-Image-SFT/blob/main/LICENSE","minGezelVersion":"1.26292.1","version":"1.0.1","releasedAt":"2026-10-19T00:00:00Z","downloadUrl":"https://huggingface.co/leejet/MOSS-Image-SFT-GGUF/resolve/238daf423f0de65316207453591f275407aac9a7/MOSS-Image-SFT-Q4_K.gguf","sha256":"902dbb0364d8b313cd91792b3cfe7c87b987386f0c5414986f3f1027b82b2225","approxSizeBytes":3327400800,"quantization":"Q4_K","recommendedSteps":50,"widthRange":{"min":512,"max":2048},"heightRange":{"min":512,"max":2048},"upstream":"https://huggingface.co/OpenMOSS-Team/MOSS-Image-SFT","category":"general","weightsKind":"diffusion-model","supportsImg2Img":false,"auxiliaryFiles":[{"role":"vae","downloadUrl":"https://huggingface.co/black-forest-labs/FLUX.2-small-decoder/resolve/a3efc24f613ef42d9428af62fdbd6f5fd8856c4a/full_encoder_small_decoder.safetensors","sha256":"ea4273f02d1fafbf8e1d1c2cf6018ed8748652eb0bf34f2dd91171f16f15ab62","approxSizeBytes":249519092,"tensors":{"floatElements":{"F32":62373347},"otherBytes":8}},{"role":"llm","downloadUrl":"https://huggingface.co/unsloth/Qwen3.5-9B-GGUF/resolve/3885219b6810b007914f3a7950a8d1b469d598a5/Qwen3.5-9B-Q4_K_M.gguf","sha256":"03b74727a860a56338e042c4420bb3f04b2fec5734175f4cb9fa853daf52b7e8","approxSizeBytes":5680522464},{"role":"tokenizer","downloadUrl":"https://huggingface.co/OpenMOSS-Team/MOSS-Image-SFT/resolve/c32bea2a87bc0d491d0412d7fad34e3a5b2ab3c4/tokenizer/tokenizer.json","sha256":"06b9509352d2af50381ab2247e083b80d32d5c0aba91c272ca9ff729b6a0e523","approxSizeBytes":19989325}],"commercialUse":true,"notes":"Requires a Gezel build with MOSS-Image support and stable-diffusion.cpp master-973 or newer (Gezel pins master-975-95b3090). Use 50 steps, CFG 3, Euler, and dimensions divisible by 16; 1 to 2 megapixels is the upstream recommendation. The engine applies MOSS's time shift of 5 and its layer-22 skip on the negative guidance branch by itself. The tokenizer is passed separately because Qwen3.5's vocabulary differs from the one the engine embeds. The VAE is the FLUX.2 small decoder, the one the upstream pipeline ships and the same file FLUX.2 Klein uses. Text-to-image only. Tested end to end in Gezel on a DGX Spark (GB10, CUDA): about 205 seconds per 1024px image at 50 steps and about 93 seconds at 768px, with a peak of about 13.5 GiB of memory. In the Gezel image bench it spelled short signs and poster titles correctly.","availableVersions":["1.0.1","1.0.0"]}},{"manifest":{"schemaVersion":1,"kind":"image-model","id":"sd-turbo","name":"SD-Turbo","description":"Stability AI's distilled SD 2.1 model that produces images in 1-4 sampler steps. Smaller and faster than SDXL Turbo, near-real-time on modest hardware. Set steps=1-4 and CFG to 1.0.","tags":["fast","distilled","low-vram","1-step"],"maintainer":{"name":"Stability AI","url":"https://huggingface.co/stabilityai/sd-turbo"},"license":"Stability AI Community License","licenseClass":"custom-restricted","licenseShortName":"Stability Community","licenseUrl":"https://huggingface.co/stabilityai/sd-turbo/blob/main/LICENSE.md","version":"1.0.1","releasedAt":"2026-10-11T00:00:00Z","downloadUrl":"https://huggingface.co/stabilityai/sd-turbo/resolve/main/sd_turbo.safetensors","sha256":"3f067a1b943cf162f2b8f8588f6cf5824bd5b4c7d1d88d87164b9ca123616549","approxSizeBytes":5214561328,"quantization":"FP16","recommendedSteps":4,"widthRange":{"min":256,"max":768},"heightRange":{"min":256,"max":768},"upstream":"https://huggingface.co/stabilityai/sd-turbo","category":"general","weightsKind":"checkpoint","supportsImg2Img":true,"auxiliaryFiles":[],"hardwareTier":"low","minRamGB":8,"commercialUse":true,"availableVersions":["1.0.1","1.0.0"]}},{"manifest":{"schemaVersion":1,"kind":"image-model","id":"sdxl-lightning-4step","name":"SDXL Lightning (4-step)","description":"ByteDance's distillation of SDXL down to 4 sampler steps with quality close to the full 30-step base. The full-unet checkpoint, not a LoRA. CFG should be 1.0. Strong choice when you want SDXL quality with near-Turbo speed and a commercially-usable license.","tags":["sdxl","fast","distilled","4-step"],"maintainer":{"name":"ByteDance","url":"https://huggingface.co/ByteDance/SDXL-Lightning"},"license":"OpenRAIL++","licenseClass":"custom-restricted","licenseShortName":"OpenRAIL++","licenseUrl":"https://huggingface.co/ByteDance/SDXL-Lightning/blob/main/LICENSE.md","version":"1.0.1","releasedAt":"2026-10-11T00:00:00Z","downloadUrl":"https://huggingface.co/ByteDance/SDXL-Lightning/resolve/main/sdxl_lightning_4step.safetensors","sha256":"e0d996ee0013e79d9d3561f50fcafb9a17e3ff07b780358e3b66d67932c4d490","approxSizeBytes":6938040682,"quantization":"FP16","recommendedSteps":4,"widthRange":{"min":512,"max":1536},"heightRange":{"min":512,"max":1536},"upstream":"https://huggingface.co/ByteDance/SDXL-Lightning","category":"general","weightsKind":"checkpoint","supportsImg2Img":true,"auxiliaryFiles":[],"hardwareTier":"mid","minRamGB":10,"commercialUse":true,"availableVersions":["1.0.1","1.0.0"]}},{"manifest":{"schemaVersion":1,"kind":"image-model","id":"sdxl-turbo","name":"SDXL Turbo","description":"Stability AI's adversarial-distilled SDXL — single-step generation at SDXL quality. Use steps=1-4, CFG=1.0. Note: license restricts commercial use to research only; for commercial work prefer SDXL Lightning.","tags":["sdxl","fast","distilled","1-step","non-commercial"],"maintainer":{"name":"Stability AI","url":"https://huggingface.co/stabilityai/sdxl-turbo"},"license":"Stability AI Non-Commercial Research","licenseClass":"commercial-restricted","licenseShortName":"Stability NC Research","licenseUrl":"https://huggingface.co/stabilityai/sdxl-turbo/blob/main/LICENSE.md","version":"1.0.1","releasedAt":"2026-10-11T00:00:00Z","downloadUrl":"https://huggingface.co/stabilityai/sdxl-turbo/resolve/main/sd_xl_turbo_1.0_fp16.safetensors","sha256":"e869ac7d6942cb327d68d5ed83a40447aadf20e0c3358d98b2cc9e270db0da26","approxSizeBytes":6938081905,"quantization":"FP16","recommendedSteps":4,"widthRange":{"min":512,"max":1024},"heightRange":{"min":512,"max":1024},"upstream":"https://huggingface.co/stabilityai/sdxl-turbo","category":"general","weightsKind":"checkpoint","supportsImg2Img":true,"auxiliaryFiles":[],"hardwareTier":"mid","minRamGB":10,"commercialUse":false,"availableVersions":["1.0.1","1.0.0"]}},{"manifest":{"schemaVersion":1,"kind":"image-model","id":"sd-1.5","name":"Stable Diffusion 1.5","description":"The classic SD 1.5 checkpoint. Tiny, fast, and runs on almost any hardware. Native 512x512; output quality is dated relative to SDXL/FLUX, but the LoRA ecosystem is massive and inference is near-instant on M-series Macs.","tags":["small","fast","low-vram","classic"],"maintainer":{"name":"Runway / Stable Diffusion community","url":"https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5"},"license":"CreativeML Open RAIL-M","licenseClass":"custom-restricted","licenseShortName":"OpenRAIL-M","licenseUrl":"https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5","version":"1.0.1","releasedAt":"2026-10-11T00:00:00Z","downloadUrl":"https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5/resolve/main/v1-5-pruned-emaonly.safetensors","sha256":"6ce0161689b3853acaa03779ec93eafe75a02f4ced659bee03f50797806fa2fa","approxSizeBytes":4265146304,"quantization":"FP16","recommendedSteps":25,"widthRange":{"min":256,"max":768},"heightRange":{"min":256,"max":768},"upstream":"https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5","category":"general","weightsKind":"checkpoint","supportsImg2Img":true,"auxiliaryFiles":[],"hardwareTier":"low","minRamGB":6,"commercialUse":true,"availableVersions":["1.0.1","1.0.0"]}},{"manifest":{"schemaVersion":1,"kind":"image-model","id":"sdxl-base-1.0","name":"Stable Diffusion XL 1.0","description":"SDXL Base — the mid-tier workhorse. Native 1024x1024 generation, strong prompt following, and a huge LoRA ecosystem. Sound default for any modern laptop with 8 GB+ memory budget.","tags":["sdxl","1024","general"],"maintainer":{"name":"Stability AI","url":"https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0"},"license":"CreativeML Open RAIL++-M","licenseClass":"custom-restricted","licenseShortName":"OpenRAIL++-M","licenseUrl":"https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/blob/main/LICENSE.md","version":"1.0.1","releasedAt":"2026-10-11T00:00:00Z","downloadUrl":"https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/resolve/main/sd_xl_base_1.0.safetensors","sha256":"31e35c80fc4829d14f90153f4c74cd59c90b779f6afe05a74cd6120b893f7e5b","approxSizeBytes":6938078334,"quantization":"FP16","recommendedSteps":30,"widthRange":{"min":512,"max":1536},"heightRange":{"min":512,"max":1536},"upstream":"https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0","category":"general","weightsKind":"checkpoint","supportsImg2Img":true,"auxiliaryFiles":[],"hardwareTier":"mid","minRamGB":10,"commercialUse":true,"availableVersions":["1.0.1","1.0.0"]}}]}
+{"schemaVersion":1,"kind":"image-model","count":14,"entries":[{"manifest":{"schemaVersion":1,"kind":"image-model","id":"animagine-xl-4","name":"Animagine XL 4.0","description":"Cagliostro Lab's flagship open anime/illustration model. SDXL-compatible so all SDXL tooling and LoRAs work. Booru-style tag prompts produce the best results. The strongest open anime model available.","tags":["sdxl","anime","illustration","artistic"],"maintainer":{"name":"Cagliostro Lab","url":"https://huggingface.co/cagliostrolab/animagine-xl-4.0"},"license":"Fair AI Public License 1.0-SD","licenseClass":"custom-restricted","licenseShortName":"Fair AI 1.0-SD","licenseUrl":"https://huggingface.co/cagliostrolab/animagine-xl-4.0","version":"4.0.1","releasedAt":"2026-10-11T00:00:00Z","downloadUrl":"https://huggingface.co/cagliostrolab/animagine-xl-4.0/resolve/main/animagine-xl-4.0.safetensors","sha256":"1d5b43ff75b6ab598502d4c779d2fbfa3dceca51c60c3b609640a60772333916","approxSizeBytes":6938434056,"quantization":"FP16","recommendedSteps":28,"widthRange":{"min":512,"max":1536},"heightRange":{"min":512,"max":1536},"upstream":"https://huggingface.co/cagliostrolab/animagine-xl-4.0","category":"artistic","weightsKind":"checkpoint","supportsImg2Img":true,"auxiliaryFiles":[],"hardwareTier":"mid","minRamGB":10,"commercialUse":true,"availableVersions":["4.0.1","4.0.0"]}},{"manifest":{"schemaVersion":1,"kind":"image-model","id":"flux-2-klein-4b-q4","name":"FLUX.2 Klein 4B (Q4)","description":"Black Forest Labs' FLUX.2-klein-4B — the successor to FLUX.1-schnell. A 4-step, CFG-free, 4B-parameter rectified flow transformer with state-of-the-art quality on consumer hardware. This package needs about 8 GB of memory and renders a 1024px image in about 7 seconds on a DGX Spark. Drops the FLUX.1 T5+CLIP encoder pair for a single Qwen3-4B LLM encoder. leejet's Q4_0 quantization of the diffusion model paired with Qwen3-4B-Q4_K_M and the FLUX.2-small-decoder VAE. Apache 2.0 throughout — fully commercial-friendly.","tags":["flux","flux2","flagship","fast","4-step","apache","llm-encoder"],"maintainer":{"name":"Black Forest Labs (quantized by leejet)","url":"https://huggingface.co/leejet/FLUX.2-klein-4B-GGUF"},"license":"Apache 2.0","licenseClass":"open","licenseShortName":"Apache 2.0","licenseUrl":"https://huggingface.co/black-forest-labs/FLUX.2-klein-4B/blob/main/LICENSE.md","recoScore":20,"version":"1.0.1","releasedAt":"2026-10-11T00:00:00Z","downloadUrl":"https://huggingface.co/leejet/FLUX.2-klein-4B-GGUF/resolve/main/flux-2-klein-4b-Q4_0.gguf","sha256":"d1023499ef3f2f82ff7c50e6778495195c1b6cc34835741778868428111f9ff4","approxSizeBytes":2460378560,"quantization":"Q4_0","recommendedSteps":4,"widthRange":{"min":512,"max":2048},"heightRange":{"min":512,"max":2048},"upstream":"https://huggingface.co/black-forest-labs/FLUX.2-klein-4B","category":"general","weightsKind":"diffusion-model","supportsImg2Img":false,"auxiliaryFiles":[{"role":"vae","downloadUrl":"https://huggingface.co/black-forest-labs/FLUX.2-small-decoder/resolve/main/full_encoder_small_decoder.safetensors","sha256":"ea4273f02d1fafbf8e1d1c2cf6018ed8748652eb0bf34f2dd91171f16f15ab62","approxSizeBytes":249519092,"tensors":{"floatElements":{"F32":62373347},"otherBytes":8}},{"role":"llm","downloadUrl":"https://huggingface.co/unsloth/Qwen3-4B-GGUF/resolve/main/Qwen3-4B-Q4_K_M.gguf","sha256":"f6f851777709861056efcdad3af01da38b31223a3ba26e61a4f8bf3a2195813a","approxSizeBytes":2497281312}],"hardwareTier":"high","minRamGB":16,"commercialUse":true,"availableVersions":["1.0.1","1.0.0"]}},{"manifest":{"schemaVersion":1,"kind":"image-model","id":"iris-3b-f32","name":"Iris 3B (F32)","description":"Generate photographs, illustrations, and artwork at about one megapixel. Iris paints pixels directly, with a separate Qwen text encoder and no VAE decoder. Approximately 14.5 GB of downloads. It renders at 50 steps, about 2.5 minutes per image on a DGX Spark.","tags":["iris","text-to-image","pixel-space","photography","illustration"],"maintainer":{"name":"Sperid Labs","url":"https://huggingface.co/speridlabs"},"license":"Apache-2.0","licenseClass":"open","licenseShortName":"Apache 2.0","licenseUrl":"https://www.apache.org/licenses/LICENSE-2.0","minGezelVersion":"1.26283","version":"1.0.2","releasedAt":"2026-10-19T00:00:00Z","downloadUrl":"https://huggingface.co/speridlabs/iris-3b/resolve/7445443349bc9abe3c96f01ff793e2098ca012b3/model.safetensors","sha256":"a90e1f4c197bac8491f82078def68884b6a3527ea7fc97a693e3e367372aa80f","approxSizeBytes":11950112680,"quantization":"F32","recommendedSteps":50,"widthRange":{"min":512,"max":1536},"heightRange":{"min":512,"max":1536},"upstream":"https://huggingface.co/speridlabs/iris-3b","category":"general","weightsKind":"diffusion-model","supportsImg2Img":false,"auxiliaryFiles":[{"role":"llm","downloadUrl":"https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct-GGUF/resolve/1cd86afb9a95c410a6038ab3b40d8b578c892266/Qwen3VL-4B-Instruct-Q4_K_M.gguf","sha256":"66358cb18bb6b3b1b6675aa412c7a88ef01d228f481184d13668e5201c730a0a","approxSizeBytes":2497281664}],"hardwareTier":"high","minRamGB":32,"commercialUse":true,"notes":"Requires a Gezel build with Iris defaults and stable-diffusion.cpp master-951-f89d9b1 or newer. Original F32 image weights are loaded in BF16; the Qwen3-VL-4B-Instruct Q4_K_M text encoder keeps its quantization. Use CFG 3, Euler, diffusion flash attention, and dimensions divisible by 16 at approximately one megapixel. The publisher recommends 100 steps; Gezel renders 50, which matched 100 in the Gezel image bench at half the time. A 20-step 1024x1024 render took 174 seconds on a 128 GB Apple M5 Max, with about 11.8 GB peak resident memory. On a DGX Spark (GB10, CUDA) a 1024px render took about 150 seconds at 50 steps and 300 at 100, with a peak of 13 to 15 GiB of memory, so 16 GB machines are tight. Captions are limited to 300 tokens including the assistant-turn suffix. No VAE or vision projector is needed. Text-to-image only: depth estimation and restoration/upscaling require separate checkpoints and are not included.","availableVersions":["1.0.2","1.0.1","1.0.0"]}},{"manifest":{"schemaVersion":1,"kind":"image-model","id":"krea-2-turbo-q4","name":"Krea 2 Turbo (Q4)","description":"Krea's flagship Krea 2 Turbo — a 12B diffusion transformer distilled for fast 8-step generation at 1K-2K resolution, paired with the Qwen-Image VAE and a Qwen3-VL text encoder. Abiray's Q4_K_M quantization is the consumer-hardware tier. Non-commercial (Krea 2 Community License).","tags":["krea","krea-2","flagship","8-step","distilled","non-commercial"],"maintainer":{"name":"Krea (quantized by Abiray)","url":"https://huggingface.co/Abiray/Krea-2-Turbo-GGUF"},"license":"Krea 2 Community License","licenseClass":"commercial-restricted","licenseShortName":"Krea 2 Community","licenseUrl":"https://huggingface.co/krea/Krea-2-Turbo/blob/main/LICENSE.pdf","version":"1.0.1","releasedAt":"2026-10-11T00:00:00Z","downloadUrl":"https://huggingface.co/Abiray/Krea-2-Turbo-GGUF/resolve/main/Krea-2-Turbo-Q4_K_M.gguf","sha256":"1e23e478144ce4c1adc6854efe2ff52930a836cf4e0008b6a0e41d235d9bdd07","approxSizeBytes":7486289232,"quantization":"Q4_K_M","recommendedSteps":8,"widthRange":{"min":512,"max":2048},"heightRange":{"min":512,"max":2048},"upstream":"https://huggingface.co/krea/Krea-2-Turbo","category":"general","weightsKind":"diffusion-model","supportsImg2Img":false,"auxiliaryFiles":[{"role":"vae","downloadUrl":"https://huggingface.co/Comfy-Org/Krea-2/resolve/main/vae/qwen_image_vae.safetensors","sha256":"a70580f0213e67967ee9c95f05bb400e8fb08307e017a924bf3441223e023d1f","approxSizeBytes":253806246,"tensors":{"floatElements":{"BF16":126892531},"otherBytes":0}},{"role":"llm","downloadUrl":"https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct-GGUF/resolve/main/Qwen3VL-4B-Instruct-Q4_K_M.gguf","sha256":"66358cb18bb6b3b1b6675aa412c7a88ef01d228f481184d13668e5201c730a0a","approxSizeBytes":2497281664}],"hardwareTier":"high","minRamGB":16,"commercialUse":false,"availableVersions":["1.0.1","1.0.0"]}},{"manifest":{"schemaVersion":1,"kind":"image-model","id":"krea-2-turbo-q6","name":"Krea 2 Turbo (Q6)","description":"Krea's flagship Krea 2 Turbo — a 12B diffusion transformer distilled for fast 8-step generation at 1K-2K resolution, paired with the Qwen-Image VAE and a Qwen3-VL text encoder. Abiray's Q6_K quantization trades extra VRAM for near-Q8 fidelity. Non-commercial (Krea 2 Community License).","tags":["krea","krea-2","flagship","8-step","distilled","non-commercial"],"maintainer":{"name":"Krea (quantized by Abiray)","url":"https://huggingface.co/Abiray/Krea-2-Turbo-GGUF"},"license":"Krea 2 Community License","licenseClass":"commercial-restricted","licenseShortName":"Krea 2 Community","licenseUrl":"https://huggingface.co/krea/Krea-2-Turbo/blob/main/LICENSE.pdf","version":"1.0.1","releasedAt":"2026-10-11T00:00:00Z","downloadUrl":"https://huggingface.co/Abiray/Krea-2-Turbo-GGUF/resolve/main/Krea-2-Turbo-Q6_K.gguf","sha256":"91103da0cf8e2b79f2108d4c6b885a62bacf7911c633373bbe0556b6da9a1f5c","approxSizeBytes":10583638350,"quantization":"Q6_K","recommendedSteps":8,"widthRange":{"min":512,"max":2048},"heightRange":{"min":512,"max":2048},"upstream":"https://huggingface.co/krea/Krea-2-Turbo","category":"general","weightsKind":"diffusion-model","supportsImg2Img":false,"auxiliaryFiles":[{"role":"vae","downloadUrl":"https://huggingface.co/Comfy-Org/Krea-2/resolve/main/vae/qwen_image_vae.safetensors","sha256":"a70580f0213e67967ee9c95f05bb400e8fb08307e017a924bf3441223e023d1f","approxSizeBytes":253806246,"tensors":{"floatElements":{"BF16":126892531},"otherBytes":0}},{"role":"llm","downloadUrl":"https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct-GGUF/resolve/main/Qwen3VL-4B-Instruct-Q4_K_M.gguf","sha256":"66358cb18bb6b3b1b6675aa412c7a88ef01d228f481184d13668e5201c730a0a","approxSizeBytes":2497281664}],"hardwareTier":"workstation","minRamGB":24,"commercialUse":false,"availableVersions":["1.0.1","1.0.0"]}},{"manifest":{"schemaVersion":1,"kind":"image-model","id":"krea-2-turbo-q8","name":"Krea 2 Turbo (Q8)","description":"Krea's flagship Krea 2 Turbo — a 12B diffusion transformer distilled for fast 8-step generation at 1K-2K resolution, paired with the Qwen-Image VAE and a Qwen3-VL text encoder. Abiray's Q8_0 quantization is near-lossless for workstation-class hardware. Non-commercial (Krea 2 Community License).","tags":["krea","krea-2","flagship","8-step","distilled","non-commercial"],"maintainer":{"name":"Krea (quantized by Abiray)","url":"https://huggingface.co/Abiray/Krea-2-Turbo-GGUF"},"license":"Krea 2 Community License","licenseClass":"commercial-restricted","licenseShortName":"Krea 2 Community","licenseUrl":"https://huggingface.co/krea/Krea-2-Turbo/blob/main/LICENSE.pdf","version":"1.0.1","releasedAt":"2026-10-11T00:00:00Z","downloadUrl":"https://huggingface.co/Abiray/Krea-2-Turbo-GGUF/resolve/main/Krea-2-Turbo-Q8_0.gguf","sha256":"8be11f7fa2b63cab28be213246001195ebe04f271a703b351461e0b775e3f775","approxSizeBytes":13705958734,"quantization":"Q8_0","recommendedSteps":8,"widthRange":{"min":512,"max":2048},"heightRange":{"min":512,"max":2048},"upstream":"https://huggingface.co/krea/Krea-2-Turbo","category":"general","weightsKind":"diffusion-model","supportsImg2Img":false,"auxiliaryFiles":[{"role":"vae","downloadUrl":"https://huggingface.co/Comfy-Org/Krea-2/resolve/main/vae/qwen_image_vae.safetensors","sha256":"a70580f0213e67967ee9c95f05bb400e8fb08307e017a924bf3441223e023d1f","approxSizeBytes":253806246,"tensors":{"floatElements":{"BF16":126892531},"otherBytes":0}},{"role":"llm","downloadUrl":"https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct-GGUF/resolve/main/Qwen3VL-4B-Instruct-Q4_K_M.gguf","sha256":"66358cb18bb6b3b1b6675aa412c7a88ef01d228f481184d13668e5201c730a0a","approxSizeBytes":2497281664}],"hardwareTier":"workstation","minRamGB":32,"commercialUse":false,"availableVersions":["1.0.1","1.0.0"]}},{"manifest":{"schemaVersion":1,"kind":"image-model","id":"ming-image-0.1-design-bf16","name":"Ming Image 0.1 Design (BF16)","description":"Generate posters, infographics, UI mockups, and text-rich visual designs locally. The 6B image transformer and separate Ling text encoder use BF16 weights. A large workstation model: approximately 49 GB of downloads; memory and speed depend on resolution and offloading.","tags":["ming","design","infographics","posters","bf16"],"maintainer":{"name":"inclusionAI (packaged by Comfy-Org)","url":"https://huggingface.co/Comfy-Org/Ming-Image"},"license":"MIT","licenseClass":"open","licenseShortName":"MIT","licenseUrl":"https://huggingface.co/inclusionAI/Ming-Image-0.1-Design/blob/main/LICENSE","minGezelVersion":"1.26283","version":"1.0.2","releasedAt":"2026-10-19T00:00:00Z","downloadUrl":"https://huggingface.co/Comfy-Org/Ming-Image/resolve/93804ff3747517a318969c5f61b25a57daefe09f/diffusion_models/ming_image_0.1_design_bf16.safetensors","sha256":"8781c6fc679f1812ad318cc7fd84c4d05a23b3571b9a7efeff73a92054ef4659","approxSizeBytes":12309858664,"quantization":"BF16","recommendedSteps":12,"widthRange":{"min":1024,"max":2048},"heightRange":{"min":1024,"max":2048},"upstream":"https://huggingface.co/inclusionAI/Ming-Image-0.1-Design","category":"general","weightsKind":"diffusion-model","supportsImg2Img":false,"auxiliaryFiles":[{"role":"vae","downloadUrl":"https://huggingface.co/Comfy-Org/Ming-Image/resolve/93804ff3747517a318969c5f61b25a57daefe09f/vae/ming_image_vae_bf16.safetensors","sha256":"7f5bed402dc8c77dc2e0ab1929a85d4df433b7cf7b599dfa8c353da98db0b90a","approxSizeBytes":253816696,"tensors":{"floatElements":{"BF16":126897716},"otherBytes":0}},{"role":"llm","downloadUrl":"https://huggingface.co/Comfy-Org/Ming-Image/resolve/93804ff3747517a318969c5f61b25a57daefe09f/text_encoders/ming_image_0.1_ling_mini_2.0_bf16.safetensors","sha256":"7eccad21934f2e6d9103b5c589a448082e89f7b5b73a6b9c6d0332be7de26c1e","approxSizeBytes":36677768245,"tensors":{"floatElements":{"BF16":18332699136,"F32":9728},"otherBytes":12210709}},{"role":"tokenizer","downloadUrl":"https://huggingface.co/inclusionAI/Ming-Image-0.1-Design/resolve/208087ada1486931692c1896f38d4cd16ff3df82/mllm/tokenizer.json","sha256":"e7ff01708d504f7bf4dbf7f5815adde57bab9a40e7f563ab6ad1acace4464917","approxSizeBytes":12210709}],"hardwareTier":"workstation","minRamGB":96,"commercialUse":true,"notes":"Requires a Gezel build with Ming support and stable-diffusion.cpp master-951-f89d9b1 or newer. Use 12 steps, CFG 1, Euler, and dimensions divisible by 16. Keep the Ling text encoder in BF16. Text-to-image only; the separate Design-Layer model is not included. BF16 avoids the missing INT8 matrix kernel observed on Metal with this engine pin. A 1024px design poster passed end-to-end testing on a 128 GB Apple M5 Max; memory guidance is conservative, not a measured minimum. Local generation took approximately 5–7 minutes with VAE tiling. On a DGX Spark (GB10, CUDA) a 1024px image took about 24 seconds, with a peak of about 46 GiB of memory. Transparent-background output is not validated on this Metal build: both a generic transparency request and the documented RGBA prefix produced opaque white backgrounds, while PNG bytes were preserved unchanged. Upstream transparency guidance: https://github.com/inclusionAI/Ming-Image#transparent-background-generation-tip.","availableVersions":["1.0.2","1.0.1","1.0.0"]}},{"manifest":{"schemaVersion":1,"kind":"image-model","id":"moss-image-sft-q4","name":"MOSS-Image SFT (Q4)","description":"OpenMOSS's MOSS-Image, fine-tuned to follow instructions. Strong at clear, accurate text inside images (signs, posters, whiteboards, labeled diagrams, Chinese as well as English), and it can place text and objects where a prompt or a JSON layout says. A 5B image transformer with Qwen3.5 9B as its text encoder and the FLUX.2 small-decoder VAE: leejet's Q4_K quantization of the transformer and a Q4_K_M encoder, about 9.3 GB of downloads. Careful rather than fast: 50 steps with guidance means about a hundred transformer passes per image. Apache 2.0 throughout, so fully commercial-friendly.","tags":["moss","text-rendering","layouts","apache","llm-encoder"],"maintainer":{"name":"OpenMOSS (quantized by leejet)","url":"https://huggingface.co/leejet/MOSS-Image-SFT-GGUF"},"license":"Apache 2.0","licenseClass":"open","licenseShortName":"Apache 2.0","licenseUrl":"https://huggingface.co/OpenMOSS-Team/MOSS-Image-SFT/blob/main/LICENSE","minGezelVersion":"1.26292.1","version":"1.0.1","releasedAt":"2026-10-19T00:00:00Z","downloadUrl":"https://huggingface.co/leejet/MOSS-Image-SFT-GGUF/resolve/238daf423f0de65316207453591f275407aac9a7/MOSS-Image-SFT-Q4_K.gguf","sha256":"902dbb0364d8b313cd91792b3cfe7c87b987386f0c5414986f3f1027b82b2225","approxSizeBytes":3327400800,"quantization":"Q4_K","recommendedSteps":50,"widthRange":{"min":512,"max":2048},"heightRange":{"min":512,"max":2048},"upstream":"https://huggingface.co/OpenMOSS-Team/MOSS-Image-SFT","category":"general","weightsKind":"diffusion-model","supportsImg2Img":false,"auxiliaryFiles":[{"role":"vae","downloadUrl":"https://huggingface.co/black-forest-labs/FLUX.2-small-decoder/resolve/a3efc24f613ef42d9428af62fdbd6f5fd8856c4a/full_encoder_small_decoder.safetensors","sha256":"ea4273f02d1fafbf8e1d1c2cf6018ed8748652eb0bf34f2dd91171f16f15ab62","approxSizeBytes":249519092,"tensors":{"floatElements":{"F32":62373347},"otherBytes":8}},{"role":"llm","downloadUrl":"https://huggingface.co/unsloth/Qwen3.5-9B-GGUF/resolve/3885219b6810b007914f3a7950a8d1b469d598a5/Qwen3.5-9B-Q4_K_M.gguf","sha256":"03b74727a860a56338e042c4420bb3f04b2fec5734175f4cb9fa853daf52b7e8","approxSizeBytes":5680522464},{"role":"tokenizer","downloadUrl":"https://huggingface.co/OpenMOSS-Team/MOSS-Image-SFT/resolve/c32bea2a87bc0d491d0412d7fad34e3a5b2ab3c4/tokenizer/tokenizer.json","sha256":"06b9509352d2af50381ab2247e083b80d32d5c0aba91c272ca9ff729b6a0e523","approxSizeBytes":19989325}],"commercialUse":true,"notes":"Requires a Gezel build with MOSS-Image support and stable-diffusion.cpp master-973 or newer (Gezel pins master-975-95b3090). Use 50 steps, CFG 3, Euler, and dimensions divisible by 16; 1 to 2 megapixels is the upstream recommendation. The engine applies MOSS's time shift of 5 and its layer-22 skip on the negative guidance branch by itself. The tokenizer is passed separately because Qwen3.5's vocabulary differs from the one the engine embeds. The VAE is the FLUX.2 small decoder, the one the upstream pipeline ships and the same file FLUX.2 Klein uses. Text-to-image only. Tested end to end in Gezel on a DGX Spark (GB10, CUDA): about 205 seconds per 1024px image at 50 steps and about 93 seconds at 768px, with a peak of about 13.5 GiB of memory. In the Gezel image bench it spelled short signs and poster titles correctly.","availableVersions":["1.0.1","1.0.0"]}},{"manifest":{"schemaVersion":1,"kind":"image-model","id":"qwen-image-edit-2511-q4","name":"Qwen Image Edit 2511 (Q4)","description":"Qwen's image editor: give it a picture and say what to change (\"make the sky a sunset,\" \"remove the car,\" \"turn it into a watercolor\") and it changes that and keeps the rest. It holds faces and characters steady across edits, rewrites text inside a picture, and combines up to three pictures into one. It only edits: it does not draw a picture from words alone. A 20B image transformer with Qwen2.5-VL 7B reading both the instruction and the picture: unsloth's Q4_K_M quantization, a Q4_K_M encoder and its vision projector, about 19.5 GB of downloads. Careful rather than fast: an edit takes about five and a half minutes on an RTX Spark N1X, and about nine when it combines two pictures. Apache 2.0 throughout, so fully commercial-friendly.","tags":["qwen","image-editing","text-rendering","apache","llm-encoder"],"maintainer":{"name":"Qwen (quantized by unsloth)","url":"https://huggingface.co/unsloth/Qwen-Image-Edit-2511-GGUF"},"license":"Apache 2.0","licenseClass":"open","licenseShortName":"Apache 2.0","licenseUrl":"https://huggingface.co/Qwen/Qwen-Image-Edit-2511/blob/main/README.md","minGezelVersion":"1.26301","version":"1.0.0","releasedAt":"2026-10-27T00:00:00Z","downloadUrl":"https://huggingface.co/unsloth/Qwen-Image-Edit-2511-GGUF/resolve/main/qwen-image-edit-2511-Q4_K_M.gguf","sha256":"8677bac90627adbbc11efab87b1870e701c4eb3689ee865a3de8ab81b705a723","approxSizeBytes":13244758624,"quantization":"Q4_K_M","recommendedSteps":20,"widthRange":{"min":512,"max":2048},"heightRange":{"min":512,"max":2048},"upstream":"https://huggingface.co/Qwen/Qwen-Image-Edit-2511","category":"general","weightsKind":"diffusion-model","supportsImg2Img":false,"auxiliaryFiles":[{"role":"vae","downloadUrl":"https://huggingface.co/Comfy-Org/Qwen-Image/resolve/main/split_files/vae/qwen_image_vae.safetensors","sha256":"a70580f0213e67967ee9c95f05bb400e8fb08307e017a924bf3441223e023d1f","approxSizeBytes":253806246,"tensors":{"floatElements":{"BF16":126892531},"otherBytes":0}},{"role":"llm","downloadUrl":"https://huggingface.co/unsloth/Qwen2.5-VL-7B-Instruct-GGUF/resolve/main/Qwen2.5-VL-7B-Instruct-Q4_K_M.gguf","sha256":"d16776dcd9a28d42758c2958ed3a752aabf20a305252cd64ff2be72b4a78c503","approxSizeBytes":4683072384},{"role":"llm_vision","downloadUrl":"https://huggingface.co/unsloth/Qwen2.5-VL-7B-Instruct-GGUF/resolve/main/mmproj-F16.gguf","sha256":"987dd0733033fb5dd9b124d1ca926ae865572e432384eee7618b2eec3e735e17","approxSizeBytes":1354163040}],"commercialUse":true,"availableVersions":["1.0.0"]}},{"manifest":{"schemaVersion":1,"kind":"image-model","id":"sd-turbo","name":"SD-Turbo","description":"Stability AI's distilled SD 2.1 model that produces images in 1-4 sampler steps. Smaller and faster than SDXL Turbo, near-real-time on modest hardware. Set steps=1-4 and CFG to 1.0.","tags":["fast","distilled","low-vram","1-step"],"maintainer":{"name":"Stability AI","url":"https://huggingface.co/stabilityai/sd-turbo"},"license":"Stability AI Community License","licenseClass":"custom-restricted","licenseShortName":"Stability Community","licenseUrl":"https://huggingface.co/stabilityai/sd-turbo/blob/main/LICENSE.md","version":"1.0.1","releasedAt":"2026-10-11T00:00:00Z","downloadUrl":"https://huggingface.co/stabilityai/sd-turbo/resolve/main/sd_turbo.safetensors","sha256":"3f067a1b943cf162f2b8f8588f6cf5824bd5b4c7d1d88d87164b9ca123616549","approxSizeBytes":5214561328,"quantization":"FP16","recommendedSteps":4,"widthRange":{"min":256,"max":768},"heightRange":{"min":256,"max":768},"upstream":"https://huggingface.co/stabilityai/sd-turbo","category":"general","weightsKind":"checkpoint","supportsImg2Img":true,"auxiliaryFiles":[],"hardwareTier":"low","minRamGB":8,"commercialUse":true,"availableVersions":["1.0.1","1.0.0"]}},{"manifest":{"schemaVersion":1,"kind":"image-model","id":"sdxl-lightning-4step","name":"SDXL Lightning (4-step)","description":"ByteDance's distillation of SDXL down to 4 sampler steps with quality close to the full 30-step base. The full-unet checkpoint, not a LoRA. CFG should be 1.0. Strong choice when you want SDXL quality with near-Turbo speed and a commercially-usable license.","tags":["sdxl","fast","distilled","4-step"],"maintainer":{"name":"ByteDance","url":"https://huggingface.co/ByteDance/SDXL-Lightning"},"license":"OpenRAIL++","licenseClass":"custom-restricted","licenseShortName":"OpenRAIL++","licenseUrl":"https://huggingface.co/ByteDance/SDXL-Lightning/blob/main/LICENSE.md","version":"1.0.1","releasedAt":"2026-10-11T00:00:00Z","downloadUrl":"https://huggingface.co/ByteDance/SDXL-Lightning/resolve/main/sdxl_lightning_4step.safetensors","sha256":"e0d996ee0013e79d9d3561f50fcafb9a17e3ff07b780358e3b66d67932c4d490","approxSizeBytes":6938040682,"quantization":"FP16","recommendedSteps":4,"widthRange":{"min":512,"max":1536},"heightRange":{"min":512,"max":1536},"upstream":"https://huggingface.co/ByteDance/SDXL-Lightning","category":"general","weightsKind":"checkpoint","supportsImg2Img":true,"auxiliaryFiles":[],"hardwareTier":"mid","minRamGB":10,"commercialUse":true,"availableVersions":["1.0.1","1.0.0"]}},{"manifest":{"schemaVersion":1,"kind":"image-model","id":"sdxl-turbo","name":"SDXL Turbo","description":"Stability AI's adversarial-distilled SDXL — single-step generation at SDXL quality. Use steps=1-4, CFG=1.0. Note: license restricts commercial use to research only; for commercial work prefer SDXL Lightning.","tags":["sdxl","fast","distilled","1-step","non-commercial"],"maintainer":{"name":"Stability AI","url":"https://huggingface.co/stabilityai/sdxl-turbo"},"license":"Stability AI Non-Commercial Research","licenseClass":"commercial-restricted","licenseShortName":"Stability NC Research","licenseUrl":"https://huggingface.co/stabilityai/sdxl-turbo/blob/main/LICENSE.md","version":"1.0.1","releasedAt":"2026-10-11T00:00:00Z","downloadUrl":"https://huggingface.co/stabilityai/sdxl-turbo/resolve/main/sd_xl_turbo_1.0_fp16.safetensors","sha256":"e869ac7d6942cb327d68d5ed83a40447aadf20e0c3358d98b2cc9e270db0da26","approxSizeBytes":6938081905,"quantization":"FP16","recommendedSteps":4,"widthRange":{"min":512,"max":1024},"heightRange":{"min":512,"max":1024},"upstream":"https://huggingface.co/stabilityai/sdxl-turbo","category":"general","weightsKind":"checkpoint","supportsImg2Img":true,"auxiliaryFiles":[],"hardwareTier":"mid","minRamGB":10,"commercialUse":false,"availableVersions":["1.0.1","1.0.0"]}},{"manifest":{"schemaVersion":1,"kind":"image-model","id":"sd-1.5","name":"Stable Diffusion 1.5","description":"The classic SD 1.5 checkpoint. Tiny, fast, and runs on almost any hardware. Native 512x512; output quality is dated relative to SDXL/FLUX, but the LoRA ecosystem is massive and inference is near-instant on M-series Macs.","tags":["small","fast","low-vram","classic"],"maintainer":{"name":"Runway / Stable Diffusion community","url":"https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5"},"license":"CreativeML Open RAIL-M","licenseClass":"custom-restricted","licenseShortName":"OpenRAIL-M","licenseUrl":"https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5","version":"1.0.1","releasedAt":"2026-10-11T00:00:00Z","downloadUrl":"https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5/resolve/main/v1-5-pruned-emaonly.safetensors","sha256":"6ce0161689b3853acaa03779ec93eafe75a02f4ced659bee03f50797806fa2fa","approxSizeBytes":4265146304,"quantization":"FP16","recommendedSteps":25,"widthRange":{"min":256,"max":768},"heightRange":{"min":256,"max":768},"upstream":"https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5","category":"general","weightsKind":"checkpoint","supportsImg2Img":true,"auxiliaryFiles":[],"hardwareTier":"low","minRamGB":6,"commercialUse":true,"availableVersions":["1.0.1","1.0.0"]}},{"manifest":{"schemaVersion":1,"kind":"image-model","id":"sdxl-base-1.0","name":"Stable Diffusion XL 1.0","description":"SDXL Base — the mid-tier workhorse. Native 1024x1024 generation, strong prompt following, and a huge LoRA ecosystem. Sound default for any modern laptop with 8 GB+ memory budget.","tags":["sdxl","1024","general"],"maintainer":{"name":"Stability AI","url":"https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0"},"license":"CreativeML Open RAIL++-M","licenseClass":"custom-restricted","licenseShortName":"OpenRAIL++-M","licenseUrl":"https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/blob/main/LICENSE.md","version":"1.0.1","releasedAt":"2026-10-11T00:00:00Z","downloadUrl":"https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/resolve/main/sd_xl_base_1.0.safetensors","sha256":"31e35c80fc4829d14f90153f4c74cd59c90b779f6afe05a74cd6120b893f7e5b","approxSizeBytes":6938078334,"quantization":"FP16","recommendedSteps":30,"widthRange":{"min":512,"max":1536},"heightRange":{"min":512,"max":1536},"upstream":"https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0","category":"general","weightsKind":"checkpoint","supportsImg2Img":true,"auxiliaryFiles":[],"hardwareTier":"mid","minRamGB":10,"commercialUse":true,"availableVersions":["1.0.1","1.0.0"]}}]}

package/dist/catalog/image-models/qw/qwen-image-edit-2511-q4/manifest.json

@@ -0,0 +1,44 @@
+{
+  "schemaVersion": 1,
+  "kind": "image-model",
+  "id": "qwen-image-edit-2511-q4",
+  "name": "Qwen Image Edit 2511 (Q4)",
+  "description": "Qwen's image editor: give it a picture and say what to change (\"make the sky a sunset,\" \"remove the car,\" \"turn it into a watercolor\") and it changes that and keeps the rest. It holds faces and characters steady across edits, rewrites text inside a picture, and combines up to three pictures into one. It only edits: it does not draw a picture from words alone. A 20B image transformer with Qwen2.5-VL 7B reading both the instruction and the picture: unsloth's Q4_K_M quantization, a Q4_K_M encoder and its vision projector, about 19.5 GB of downloads. Careful rather than fast: an edit takes about five and a half minutes on an RTX Spark N1X, and about nine when it combines two pictures. Apache 2.0 throughout, so fully commercial-friendly.",
+  "tags": [
+    "qwen",
+    "image-editing",
+    "text-rendering",
+    "apache",
+    "llm-encoder"
+  ],
+  "maintainer": {
+    "name": "Qwen (quantized by unsloth)",
+    "url": "https://huggingface.co/unsloth/Qwen-Image-Edit-2511-GGUF"
+  },
+  "license": "Apache 2.0",
+  "licenseClass": "open",
+  "licenseShortName": "Apache 2.0",
+  "licenseUrl": "https://huggingface.co/Qwen/Qwen-Image-Edit-2511/blob/main/README.md",
+  "minGezelVersion": "1.26301",
+  "yankedVersions": [],
+  "category": "general",
+  "recommendedSteps": 20,
+  "widthRange": {
+    "min": 512,
+    "max": 2048
+  },
+  "heightRange": {
+    "min": 512,
+    "max": 2048
+  },
+  "upstream": "https://huggingface.co/Qwen/Qwen-Image-Edit-2511",
+  "weightsKind": "diffusion-model",
+  "promptStyle": "sentences",
+  "rendersText": true,
+  "supportsImg2Img": false,
+  "referenceImages": {
+    "max": 3
+  },
+  "editOnly": true,
+  "commercialUse": true
+}

package/dist/catalog/image-models/qw/qwen-image-edit-2511-q4/versions/1.0.0/manifest.json

@@ -0,0 +1,36 @@
+{
+  "schemaVersion": 1,
+  "version": "1.0.0",
+  "releasedAt": "2026-10-27T00:00:00Z",
+  "minGezelVersion": "1.26301",
+  "downloadUrl": "https://huggingface.co/unsloth/Qwen-Image-Edit-2511-GGUF/resolve/main/qwen-image-edit-2511-Q4_K_M.gguf",
+  "sha256": "8677bac90627adbbc11efab87b1870e701c4eb3689ee865a3de8ab81b705a723",
+  "approxSizeBytes": 13244758624,
+  "quantization": "Q4_K_M",
+  "auxiliaryFiles": [
+    {
+      "role": "vae",
+      "downloadUrl": "https://huggingface.co/Comfy-Org/Qwen-Image/resolve/main/split_files/vae/qwen_image_vae.safetensors",
+      "sha256": "a70580f0213e67967ee9c95f05bb400e8fb08307e017a924bf3441223e023d1f",
+      "approxSizeBytes": 253806246,
+      "tensors": {
+        "floatElements": {
+          "BF16": 126892531
+        },
+        "otherBytes": 0
+      }
+    },
+    {
+      "role": "llm",
+      "downloadUrl": "https://huggingface.co/unsloth/Qwen2.5-VL-7B-Instruct-GGUF/resolve/main/Qwen2.5-VL-7B-Instruct-Q4_K_M.gguf",
+      "sha256": "d16776dcd9a28d42758c2958ed3a752aabf20a305252cd64ff2be72b4a78c503",
+      "approxSizeBytes": 4683072384
+    },
+    {
+      "role": "llm_vision",
+      "downloadUrl": "https://huggingface.co/unsloth/Qwen2.5-VL-7B-Instruct-GGUF/resolve/main/mmproj-F16.gguf",
+      "sha256": "987dd0733033fb5dd9b124d1ca926ae865572e432384eee7618b2eec3e735e17",
+      "approxSizeBytes": 1354163040
+    }
+  ]
+}