@melite/models-fibo2-edit 0.2.0-alpha → 0.2.2-alpha

package/README.md

@@ -1,30 +1,39 @@
----
-description: Fibo 2.0 Edit — BRIA's image-editing model (the Comfy-Org repack); the third image-edit family, encoder-emitted latent, dev/edit-turbo lanes.
-kind: package-library
----
-
 # fibo2-edit
 
-fibo2-edit registers Fibo 2.0 Edit (form `form_edit`, label "Fibo 2.0 Edit", genre `image`): BRIA's licensed image-editing model, boarded from the Comfy-Org repack as the image-edit card's third family beside the two qwen-img-edit lanes. The lane: UNETLoader → CLIPLoader (combo `fibo2`) + VAELoader → TextEncodeFibo2 — ONE encoder call emits positive AND negative (an empty `negative_prompt` rides the model's trained null caption, so CFG works text-only), cites input images in-prompt as `<image_N>` (up to 16), and — with the VAE seat connected — splices reference latents into BOTH conditionings → KSampler on the encoder's own emitted latent (128-channel, trained size: the node's width/height stay 0 = the first input image's aspect at ~1MP) → decode and save under `<run>/edit`. With `input_mask` + `mask_mode: crop_stitch` the mask rides MaskToSEGS → DetailerForEach instead (no KSampler, no decode under it).
+fibo2-edit registers Fibo 2.0 Edit (form `form_edit`, label "Fibo 2.0 Edit", genre `image`): BRIA's licensed image-editing model, boarded from the Comfy-Org repack as the image-edit card's third family beside the two qwen-img-edit lanes. The lane: UNETLoader → CLIPLoader (combo `fibo2`) + VAELoader → TextEncodeFibo2 — ONE encoder call emits positive AND negative (an empty `negative_prompt` rides the trained null caption — CFG works text-only), cites input images in-prompt as `<image_N>` (up to 16), and — with the VAE seat connected — splices reference latents into BOTH conditionings → KSampler on the encoder's own emitted latent (128-channel; the node's zero sizes mean the first input's aspect at ~1MP) → decode and save under `<run>/edit`. With `input_mask` + `mask_mode: crop_stitch` the mask rides MaskToSEGS → DetailerForEach instead (no KSampler, no decode under it).
 
 ## The BF16-encoder rule (upstream issue 17248, open)
 
-The INT8 Qwen3-VL text encoder fails on image-conditioned encoding. Every edit lane here encodes through the BF16 encoder (`qwen3vl_4b_bf16.safetensors`, provisioned); the INT8 encoder's seat is the text-only FLUX-shared use, and it is deliberately NOT provisioned — the manage rail never offers it for this family.
+The INT8 Qwen3-VL text encoder fails on image-conditioned encoding. Every edit lane encodes through the BF16 encoder (`qwen3vl_4b_bf16.safetensors`, provisioned); the INT8 encoder's seat is the text-only FLUX-shared use, deliberately NOT provisioned — the manage rail never offers it for this family.
 
 ## Params and law
 
 Declared surface: `input_images`, `image1`, `image2`, `image3`, `reference_images`, `input_mask` (validation-only aliases the lane admits as wire keys), `prompt`, `negative_prompt`, `match_input_resolution` (the declared no-op, draft compatibility), `seed`, `steps`, `cfg`, `sampler`, `scheduler`. Laws: `prompt` is required (the instruction, not a description of the image); an anchor is required — the ordered `input_images` list or a legacy spelling; the image cap is 16 total, anchor included (`max_images`); no stated size exists (the encoder emits the latent — a draft stating width/height refuses field-named); the solver span stays steps 1–50. Defaults: 30 steps, `euler` / `simple`, cfg 5 — the dev lane's official pair.
 
-Variants: `dev` (30 steps CFG 5, the reference checkpoint) and `edit-turbo` (8 steps CFG 1, the distilled checkpoint), each with `int8_convrot` and `bf16` quant cells (8.4 GB / 16.5 GB unets). Every variant declares its `sampler_envelope`; the community 4-step turbo report is deliberately not pinned. NO Recipe group: nothing upstream canonical enough to pin, and a stamp-less Recipe group is exactly what recipe-rows LAW A refuses — a canned one-click lands with upstream truth.
+Variants: `dev` (30 steps CFG 5, the reference checkpoint) and `edit-turbo` (8 steps CFG 1, the distilled checkpoint), each with `int8_convrot` and `bf16` quant cells (8.4 GB / 16.5 GB unets). Every variant declares its `sampler_envelope`; the community 4-step turbo report is deliberately not pinned. NO Recipe group — nothing upstream is canonical enough to pin, and a stamp-less Recipe group is what recipe-rows LAW A refuses: a canned one-click lands with upstream truth.
 
 ## Mod surface
 
-The DECL boards `mask` and `loras`. `./mods/mask` is the bare flavor whose region contract rides the crop_stitch detailer over the encoder conditioning; automatic inpainting with a named field refuses on every lane — no lane inpaints in place. `./mods/loras` is the seam with an empty row set (no first-party Fibo adapters); user files ride the same bag the pack registers.
+The DECL boards `mask` and `loras`. `./mods/mask` is the bare flavor whose region contract rides the crop_stitch detailer over the encoder conditioning; automatic inpainting with a named field refuses on every lane — no lane inpaints in place. `./mods/loras` is the seam with an empty row set (no first-party Fibo adapters); user files ride the pack's bag.
+
+## Model Experience
+
+#### What the model sees
+
+TextEncodeFibo2 receives the instruction, negative, and the anchor images' spliced reference latents; KSampler receives the emitted latent and solver values.
+
+#### Token effect
+
+Zero-direct — the instruction rides the composed graph, never language-model context.
+
+#### KV Cache effect
+
+Independent — each edit is its own engine request.
 
 ## Known Limitations and Deferred Work
 
 - **INT8 encoder unusable for edits** — upstream issue 17248 (open): the int8 vision tower fails on image-conditioned encoding; BF16 encoder only, pinned in `lane_inputs.text_encoder`.
-- **Graph goldens deferred** — the golden driver's wiring gate judges against the recorded boot fixture (`tests/walls/fixtures/object_info.boot.json`), which predates `TextEncodeFibo2`; goldens land the day the engine hop re-records it. The branch pins ride `tests/action.spec.ts` structural asserts meanwhile.
+- **Graph goldens deferred** — the golden driver's wiring gate judges against the recorded boot fixture (`tests/walls/fixtures/object_info.boot.json`), which predates `TextEncodeFibo2`; goldens land the day the engine re-records it. The branch pins ride `tests/action.spec.ts` structural asserts meanwhile.
 - **Facts rows pending the pin audit** — `model-facts.json` (the `fibo2-edit` card row + `core_node_types` entry for `comfy_extras/nodes_fibo2.py`) and the recorded engine's `engine_node_types` are the engine pin audit's writes; until they land the offered-cells and flows walls carry named pending reds.
 - **Live behavior unobserved** — not verified against the running model.
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@melite/models-fibo2-edit",
-  "version": "0.2.0-alpha",
+  "version": "0.2.2-alpha",
   "license": "AGPL-3.0-or-later",
   "type": "module",
   "exports": {
@@ -19,12 +19,7 @@
     }
   },
   "dependencies": {
-    "@melite/comfyui-card-kit": "0.2.0-alpha"
-  },
-  "devDependencies": {
-    "@types/node": "^22.20.0",
-    "typescript": "^5.9.3",
-    "@deepseek-ai/cordis": "^4.0.5"
+    "@melite/comfyui-card-kit": "0.2.2-alpha"
   },
   "files": [
     "lib/index.js",
@@ -34,10 +29,27 @@
   "peerDependencies": {
     "@deepseek-ai/cordis": "^4.0.5"
   },
+  "description": "Fibo 2.0 Edit — BRIA's image-editing model (the Comfy-Org repack); the third image-edit family, encoder-emitted latent, dev/edit-turbo lanes.",
+  "repository": {
+    "type": "git",
+    "url": "https://github.com/JayDataEngineer/melite.git",
+    "directory": "packages/models/fibo2-edit"
+  },
+  "keywords": [
+    "melite",
+    "comfyui",
+    "cordis",
+    "agent",
+    "media-generation",
+    "library"
+  ],
   "scripts": {
     "build": "tsc -p tsconfig.build-mods.json && tsc -p tsconfig.build.json",
     "typecheck": "tsc --noEmit"
   },
+  "engines": {
+    "node": ">=22.18.0"
+  },
   "publishConfig": {
     "access": "public"
   }