@shop-os/modelkit 0.7.8 → 0.7.11

package/CHANGELOG.md

@@ -9,6 +9,28 @@ header in the release PR.
 
 ## [Unreleased]
 
+## [0.7.11] - Unreleased
+
+Kling 4.0 on fal.
+
+### Added
+- `kling-4.0` (fal `kling/v4`): one row for text-, image- and reference-to-video. Reference images or videos go to `kling/v4/reference-to-video` (`image_urls`, `video_urls`, and a start frame as `start_image_url`), else a start frame to `image-to-video` (with an optional end frame), else `text-to-video`. 720p / 1080p, 3-30 s (whole seconds), aspect ratios `adaptive` (fal's `auto` on reference-to-video), 16:9, 9:16, 1:1, 21:9. References: 10 images (the start frame counts toward them) and 5 videos, each 3-30.05 s and 30 s combined. An end frame with references, more than 10 images with the start frame, and a blank prompt are refused before the call.
+- `kling-4.0-edit` (fal `kling/v4/video-to-video/edit`): a 3-30.05 s source plus up to 10 reference images, 720p / 1080p. `durationSeconds` 3-30 sets the output length; left out, the output (and its usage) is the source's length.
+- Prompts name inputs `@Image1`, `@Video1` (no space); on an edit the source is `@Video1`. Kit requests send no Kling elements.
+- Neither endpoint has a sound field, so both rows declare `audio: "always"` and drop `generateAudio` with an `option_ignored` warning.
+- Usage reports the output seconds and resolution. fal bills a request with reference videos, and every edit, on input plus output seconds (at least twice the output) at a lower rate; pricing that is the consumer's job (pro-backend's video-input rate).
+
+## [0.7.10] - Unreleased
+
+Kling V3 Omni moves from Replicate to fal at half the provider cost.
+
+### Changed
+- `kling-v3-omni-video` runs on fal Kling O3 (`adapter: "fal"`, `providerModelId: "fal-ai/kling-video/o3"`) instead of Replicate `kwaivgi/kling-v3-omni-video`. The id, aliases, ratios, resolutions, durations, defaults and the `mode` to resolution rule stay. The tier picks the endpoint path (`standard` 720p, `pro` 1080p, `4k`), and the inputs pick the task: reference images go to `reference-to-video`, frames alone to `image-to-video`, a prompt alone to `text-to-video`. A lone end frame still goes as the start frame. fal cost per second: standard $0.084 silent / $0.112 sound, pro $0.112 / $0.14, 4k $0.42 (Replicate: $0.168 / $0.224, $0.224 / $0.28, $0.42).
+- The row's usage unit is `seconds` (was Replicate `compute_seconds`): usage events now carry provider `fal`, the requested length and the rendered resolution.
+- The row declares `videos: 0, audios: 0` (was no video slot and a dormant `imagesWithVideo: 4`). The Replicate adapter never sent a reference video, so nothing that worked is lost; the resolver now trims one with `references_trimmed` instead of passing it through. Frames and reference images share fal's cap of 7, as on `kling-o3`. The display description no longer claims editing or reference video.
+- The Replicate `kwaivgi/kling-v3-omni-video` mapper stays. `seed-model-catalog` never updates an existing row, so each deployed `model_catalog` row keeps running on Replicate after the bump, until it is switched to this row's `adapter`, `providerModelId`, `references` and `usage` (pro-admin, per env). Reverting that edit rolls back.
+- Consumers that branch on `adapter === "replicate"` for this row (shopos-ui vibe-triggers, the trigger-workflows wire test) must handle the fal row before an env is switched.
+
 ## [0.7.8] - Unreleased
 
 ENG-3324: Ray 3.2 offers 10 s clips.

package/dist/catalog.cjs

@@ -4778,6 +4778,8 @@ function range(min, max) {
 }
 var KLING_PROMPT = { referenceSyntax: "numbered", maxChars: 2500 };
 var H3_PROMPT = { referenceSyntax: "numbered", maxChars: 5e4 };
+var KLING_V4_PROMPT = { referenceSyntax: "numbered", maxChars: 2500 };
+var KLING_V4_AUDIO = "always";
 var NO_REFERENCES = { images: 0, videos: 0, audios: 0 };
 function klingMotionControl(tier) {
   const label = tier === "pro" ? "Pro" : "Standard";
@@ -4878,6 +4880,79 @@ var VIDEO_FAMILY_MODELS = [
     },
     capabilities: { acceptsSourceVideo: true, videoEdit: true }
   },
+  {
+    // https://fal.ai/models/kling/v4/reference-to-video/api (+ /text-to-video,
+    // /image-to-video): one row; references (images or videos) go to reference-to-video,
+    // else a start frame to image-to-video, else text-to-video. Reference images: 10, the
+    // start frame included (fal: "counts toward the image reference limit"), 300 px min.
+    // Reference videos: 5, 3-30.05 s each and 30 s combined, 700 px min, 23.9-60.05 fps.
+    // The prompt names references as @Image1, @Video1 (no space; @Element1 for elements,
+    // which the kit does not send). No sound switch on any endpoint.
+    id: "kling-4.0",
+    aliases: ["kling-4", "kling-v4"],
+    label: "Kling 4.0",
+    type: "video",
+    adapter: "fal",
+    providerModelId: "kling/v4",
+    lifecycle: "active",
+    exposure: ["cowork", "space_builder"],
+    constraints: {
+      aspectRatios: ["adaptive", "16:9", "9:16", "1:1", "21:9"],
+      resolutions: ["720p", "1080p"],
+      durations: { min: 3, max: 30 },
+      durationPresets: [5, 10, 15, 30],
+      references: {
+        ...NO_REFERENCES,
+        images: 10,
+        videos: 5,
+        videoSeconds: { min: 3, max: 30.05, total: 30 }
+      },
+      supports: ["lastFrame"],
+      prompt: KLING_V4_PROMPT,
+      audio: KLING_V4_AUDIO
+    },
+    defaults: { durationSeconds: 5, resolution: "1080p", aspectRatio: "16:9" },
+    usage: { ...VIDEO_USAGE },
+    pricingKind: "video_per_second",
+    display: {
+      provider: "kuaishou",
+      logo: KLING_LOGO,
+      description: "Kuaishou Kling 4.0: text-to-video, start/end frames, or reference-to-video from up to 10 images and 5 videos (30 seconds combined), 3-30 seconds at 720p or 1080p. Prompts name references @Image1, @Video1."
+    },
+    capabilities: { lastFrameRequiresFirstFrame: true }
+  },
+  {
+    // https://fal.ai/models/kling/v4/video-to-video/edit/api: source 3-30.05 s (700 px
+    // min, 23.9-60.05 fps), up to 10 reference images. duration 3-30, or left out for the
+    // source's length. The prompt names the source @Video1 and images @Image1, @Image2.
+    id: "kling-4.0-edit",
+    aliases: ["kling-4-edit", "kling-v4-edit"],
+    label: "Kling 4.0 Edit",
+    type: "video",
+    adapter: "fal",
+    providerModelId: "kling/v4/video-to-video/edit",
+    lifecycle: "active",
+    exposure: ["cowork", "space_builder"],
+    constraints: {
+      resolutions: ["720p", "1080p"],
+      durations: { min: 3, max: 30 },
+      references: { ...NO_REFERENCES, images: 10 },
+      requires: ["sourceVideo"],
+      sourceVideoSeconds: { min: 3, max: 30.05 },
+      supports: [],
+      prompt: KLING_V4_PROMPT,
+      audio: KLING_V4_AUDIO
+    },
+    defaults: { resolution: "1080p" },
+    usage: { ...VIDEO_USAGE },
+    pricingKind: "video_per_second",
+    display: {
+      provider: "kuaishou",
+      logo: KLING_LOGO,
+      description: "Kuaishou Kling 4.0 video edit: change a 3-30 second clip with a prompt and up to 10 reference images, at 720p or 1080p, as long as the source or a set 3-30 seconds. Prompts name the source @Video1 and images @Image1."
+    },
+    capabilities: { acceptsSourceVideo: true, videoEdit: true }
+  },
   {
     // https://fal.ai/models/minimax/h3-max-turbo/text-to-video/api (+ /image-to-video).
     // image-to-video takes a start frame, an end frame, or both, and no aspect_ratio.
@@ -5736,38 +5811,38 @@ var VIDEO_MODELS = [
     capabilities: { ...FRAMES_EXCLUDE_REFERENCES }
   },
   {
-    // mode standard/pro/4k selects 720p/1080p/4K; the endpoint takes no separate resolution.
-    // Replicate: "Max 7 [images] without video, 4 with video"; 4K does not support
-    // reference_video. No `videos` limit yet: the adapter sends no reference video,
-    // so `imagesWithVideo` stays dormant until one is wired through.
+    // fal Kling O3 (adapters/fal-kling-omni.ts): mode standard/pro/4k selects 720p/1080p/4K
+    // in the endpoint path. Moved off Replicate kwaivgi/kling-v3-omni-video, which fal sells
+    // at half the price; the id and inputs stay. The Replicate mapper stays for deployed
+    // model_catalog rows not yet switched. Frames and reference images share fal's cap of 7.
+    // Reference video, multi-shot and keep-sound were never wired, so stay off.
     id: "kling-v3-omni-video",
     aliases: ["kwaivgi/kling-v3-omni-video", "kling-v3-omni"],
     label: "Kling V3 Omni",
     type: "video",
-    adapter: "replicate",
-    providerModelId: "kwaivgi/kling-v3-omni-video",
+    adapter: "fal",
+    providerModelId: "fal-ai/kling-video/o3",
     lifecycle: "active",
     exposure: ["cowork", "space_builder"],
     constraints: {
       aspectRatios: [...KLING_OMNI_RATIOS],
       resolutions: ["720p", "1080p", "4k"],
       durations: { min: 3, max: 15 },
-      references: { images: 7, imagesWithVideo: 4 },
+      references: { images: 7, videos: 0, audios: 0 },
       supports: ["generateAudio", "lastFrame"],
       prompt: KLING_PROMPT2,
       audio: "optional"
     },
     defaults: { durationSeconds: 5, aspectRatio: "16:9", generateAudio: true, mode: "pro" },
-    usage: { unit: "compute_seconds" },
+    usage: { ...VIDEO_USAGE3 },
     pricingKind: "video_per_second",
     display: {
       provider: "kuaishou",
       logo: "https://cdn.jsdelivr.net/npm/@lobehub/icons-static-png@1.97.0/light/kling-color.png",
-      description: "Kuaishou Kling Video 3.0 Omni unified generation and editing from text, start/end images, up to 7 reference images, and reference video; up to 15 seconds at up to 4K."
+      description: "Kuaishou Kling Video 3.0 Omni generation from text, start/end images and up to 7 reference images, with optional native audio; up to 15 seconds at up to 4K."
     }
-    // end_image "Requires start_image", but no lastFrameRequiresFirstFrame: SpaceLab resolves
-    // replicate rows in error mode, so the flag would refuse its end-frame-only nodes, which
-    // the adapter sends as start_image.
+    // No lastFrameRequiresFirstFrame: fal image-to-video needs a start frame, so the adapter
+    // sends a lone end frame as the start frame, as the Replicate model took it.
   },
   {
     id: "minimax-h3",
@@ -8522,6 +8597,7 @@ var KLING_TIERS = [
   { mode: "pro", resolution: "1080p" },
   { mode: "4k", resolution: "4k" }
 ];
+var KLING_O3_OMNI = "fal-ai/kling-video/o3";
 function klingTier(req) {
   return KLING_TIERS.find((tier) => tier.resolution === req.resolution) ?? KLING_TIERS.find((tier) => tier.mode === req.mode) ?? KLING_TIERS[0];
 }
@@ -8531,6 +8607,8 @@ var MOTION_PRO = "fal-ai/kling-video/v3/pro/motion-control";
 var MOTION_STANDARD = "fal-ai/kling-video/v3/standard/motion-control";
 var O3_REFERENCE = "fal-ai/kling-video/o3/pro/reference-to-video";
 var O3_EDIT = "fal-ai/kling-video/o3/pro/video-to-video/edit";
+var KLING_V4 = "kling/v4";
+var KLING_V4_EDIT = "kling/v4/video-to-video/edit";
 var H3_TURBO = "minimax/h3-max-turbo";
 var H3_RECAST = "minimax/h3-max/recast";
 var H3_RELIGHT = "minimax/h3-max/relight";
@@ -8539,6 +8617,7 @@ var HAPPY_HORSE_EDIT = "alibaba/happy-horse/video-edit";
 var VIDU_Q4 = "fal-ai/vidu/q4";
 var OMNIHUMAN_V2 = "bytedance/omnihuman/v2";
 var KLING_O3_IMAGE_CAP = 7;
+var KLING_V4_IMAGE_CAP = 10;
 var TURBO_AUDIO_MIN_SECONDS = 2;
 var OMNIHUMAN_AUDIO_SECONDS = { min: 4, max: 300 };
 function missing(message) {
@@ -8594,6 +8673,23 @@ function o3Reference(req) {
   }
   return needsPrompt(req);
 }
+function hasReferences(req) {
+  return count(req.referenceImages) + count(req.referenceVideos) > 0;
+}
+function klingV4(req) {
+  if (hasReferences(req)) {
+    if (req.lastFrame !== void 0) {
+      return unsupported2("takes an end frame only without reference images or videos");
+    }
+    const images = count(req.referenceImages) + (req.firstFrame === void 0 ? 0 : 1);
+    if (images > KLING_V4_IMAGE_CAP) {
+      return unsupported2(
+        `takes up to ${KLING_V4_IMAGE_CAP} images in all, the start frame included; this request has ${images}`
+      );
+    }
+  }
+  return needsPrompt(req);
+}
 function h3Turbo(req) {
   if (count(req.referenceImages) + count(req.referenceVideos) > 0) {
     return unsupported2("takes no reference images or videos");
@@ -8670,7 +8766,10 @@ var RULES = /* @__PURE__ */ new Map([
   [MOTION_PRO, motionControl],
   [MOTION_STANDARD, motionControl],
   [O3_REFERENCE, o3Reference],
+  [KLING_O3_OMNI, o3Reference],
   [O3_EDIT, (req) => needsSource(req) ?? needsPrompt(req)],
+  [KLING_V4, klingV4],
+  [KLING_V4_EDIT, (req) => needsSource(req) ?? needsPrompt(req)],
   [H3_TURBO, h3Turbo],
   [MINIMAX_H3_INSERT, h3Insert],
   [H3_RECAST, h3Recast],
@@ -8699,13 +8798,14 @@ function h3EndFrameRatioWarning(model, ratio) {
     message: `${model.id} follows the end frame's aspect ratio`
   };
 }
-function isReplicateKling(model) {
+function takesKlingMode(model) {
+  if (model.adapter === "fal") return model.providerModelId.toLowerCase() === KLING_O3_OMNI;
   if (model.adapter !== "replicate") return false;
   const kind = videoKind(model.providerModelId);
   return kind === "kling" || kind === "kling-omni";
 }
-function fillReplicateKlingResolution(model, req) {
-  if (isReplicateKling(model) && req.resolution === void 0) {
+function fillKlingModeResolution(model, req) {
+  if (takesKlingMode(model) && req.resolution === void 0) {
     req.resolution = klingTier(req).resolution;
   }
 }
@@ -8719,6 +8819,19 @@ function applyFalVideoRules(model, req, callerRatio, warnings) {
     warnings.push(viduSilentWarning(model));
     req.generateAudio = false;
   }
+  if (id === KLING_V4 && req.firstFrame !== void 0 && !hasReferences(req)) {
+    if (req.aspectRatio !== void 0) {
+      if (callerRatio !== void 0) {
+        warnings.push({
+          code: "ratio_dropped",
+          field: "aspectRatio",
+          from: req.aspectRatio,
+          message: `${model.id} follows the start frame's aspect ratio`
+        });
+      }
+      delete req.aspectRatio;
+    }
+  }
   const endFrameOnly = req.lastFrame !== void 0 && req.firstFrame === void 0;
   if (id === H3_TURBO && endFrameOnly && req.aspectRatio !== void 0) {
     if (callerRatio !== void 0 && callerRatio !== "adaptive") {
@@ -9098,10 +9211,10 @@ function resolveVideoOptions(model, request, onUnsupported = "error") {
     warnings
   );
   if (aspectRatio !== void 0) merged.aspectRatio = aspectRatio;
-  fillReplicateKlingResolution(model, merged);
+  fillKlingModeResolution(model, merged);
   const resolution = resolveVideoResolution(merged.resolution, model, onUnsupported, warnings);
   if (resolution !== void 0) merged.resolution = resolution;
-  if (isReplicateKling(model)) delete merged.mode;
+  if (takesKlingMode(model)) delete merged.mode;
   const duration = resolveDuration(
     merged.durationSeconds,
     c.durations,