Repository navigation
fix(website): tell users Seedance blocks realistic faces - #19638
Conversation
Seedance 2.5 and 2.0 refuse input images and videos that show a realistic human face, AI-generated ones included, but the Router reports that only as a generic content policy block. Blocked takes on those operations, in Cinematic Studio and on their model pages, now say so and suggest an image without a realistic face or a stylized character.
|
The latest updates on your projects. Learn more about Vercel for GitHub. 2 Skipped Deployments
|
🧩 Custom Nodes Ecosystem Matrix: 🟡 Withheld — could not measureExecutes the frontend JS of every registry pack that ships any, against this change, in both renderers. The eight This candidate changes nothing the matrix measures, so every expensive stage was skipped. Reproduce locally# build one shard's fixture, then run it
python3 scripts/registry-census/build_matrix.py --shard 1/4
MATRIX_RENDERER=vue MATRIX_VUE=1 \
npx vitest run --config vitest.matrix.config.mts
Full docs: scripts/registry-census/README.md 🌐 Website E2ETip All tests passed.
🌾 Fallow: ✅ PassedFallow audits the diff for duplication, complexity and dead code. It gates on new findings only, so when the audit completes, a red is something this PR introduced or moved — never inherited backlog. A red can also mean the audit itself did not finish; the details below say which happened. Fallow gates on new findings in changed files only, so everything below was introduced or moved by this PR (14 changed files).
Pre-existing findings are baselined in Reproduce locally# the base the gate compared against
npx fallow audit --root . --quiet --format json \
--changed-since 4f7dee93d04776604b6d1edf5614eb9574b53b75
Full docs: ADR-DEVEX-LINT-0015 🔗 Website PreviewWebsite Preview: https://comfy-website-preview-pr-19638.vercel.app Models is built and gated by PostHog (workshop-enabled), using staging Cloud. This commit: https://website-frontend-ok4f05axu-comfyui.vercel.app Last updated: 2026-09-30T17:23:58Z for |
|
Navigate logical layers of code changes, visualize relationships, and explore their blast radius. No actionable comments were generated in the recent review. 🎉 ℹ️ Recent review info⚙️ Run configurationConfiguration used: Repository: Comfy-Org/ComfyUI_frontend/.coderabbit.yaml Review profile: ASSERTIVE Plan: Advanced Run ID: 📒 Files selected for processing (1)
Included review availability: This review used your included allowance. Your plan provides up to 8 included reviews per hour; 6 remain after this review. 📝 WalkthroughWalkthroughThe workshop now identifies model slugs with a documented realistic-human-face restriction and displays a localized policy message for matching policy failures. Cinematic Studio carries the failed run’s model slug into take state and uses it to select that message. Other failures retain the generic message. ChangesProvider-specific policy messages
Priority: ⬇️ Low Estimated code review effort: 3 (Moderate) | ~20 minutes Change: Bug fix Sequence Diagram(s)sequenceDiagram
participant renderTake
participant takeFailure
participant ReelReducer
participant CinematicTakeNotice
participant refusesRealFaces
renderTake->>takeFailure: Pass model.slug with the failure
takeFailure->>ReelReducer: Dispatch takeFailed with runSlug
ReelReducer->>CinematicTakeNotice: Store failed take with runSlug
CinematicTakeNotice->>refusesRealFaces: Check failed take runSlug
refusesRealFaces-->>CinematicTakeNotice: Return whether slug matches
Merge Risk: ⚪ Minimal · up to The guidance suggests checking the prompt as well as the image, reducing the chance users are directed only toward changing face content. No concrete merge-blocking risk is established. Architecture SummaryArchitecture risk: 🔵 Low · up to The change affects 1 system. Changed systems: Architecture concerns Review detailsSystems and components
Before / after behavior
🚥 Pre-merge checks | ✅ 7✅ Passed checks (7 passed)
✨ Finishing Touches📝 Generate docstrings
🧪 Generate unit tests (beta)
Comment |
This comment has been minimized.
This comment has been minimized.
Codecov Report❌ Patch coverage is
@@ Coverage Diff @@
## main #19638 +/- ##
==========================================
- Coverage 84.46% 84.36% -0.11%
==========================================
Files 3341 3269 -72
Lines 161168 158997 -2171
Branches 41915 41195 -720
==========================================
- Hits 136130 134134 -1996
+ Misses 23142 23076 -66
+ Partials 1896 1787 -109
Flags with carried forward coverage won't be shown. Click here to find out more.
... and 72 files with indirect coverage changes 🚀 New features to boost your workflow:
|
benjcooley
left a comment
There was a problem hiding this comment.
👨💻 BILL: APPROVED at 5039ba2590, a full review of the diff against main.
| Severity | Count |
|---|---|
| BLOCKER | 0 |
| SHOULD FIX | 0 |
| TRIVIAL | 1 |
- One rule, two callers:
REAL_FACE_REFUSALSinworkshop-model-restrictions.tsfollows the file's existing restriction pattern (source + reviewed date). All 7 slugs exist in the catalogue, and the 4 Seedance 2.x pages left out are exactly the text-to-video ones, which take no image or video input. - Model page:
ModelDetailpassespolicyMessageonly for those pages, andPlaygroundOutputuses it only whenreason === 'policy'. The no-credits and member paths still come first. - Cinematic Studio:
takeFailurerecordsmodel.slug, which is the operation that actually ran.loadModel(slug)is per take, so a first-frame take is checked against the first/last-frame page, not the directed model.runSlugis threaded throughReelEvent→reduceReel, andCinematicTakeNoticeuses the rule only forblockedfailed takes. The model-load-failure path has norunSlugand correctly keeps the generic text. - Copy: en and zh-CN are both present.
Verified: restrictions, PlaygroundOutput, ModelDetail, cinematic-studio components/lib and useCinematicStudioRun pass 566/566 (43 files), typecheck clean, clean merge with current main. All required CI is green.
The Router drops BytePlus's face code, so on the Seedance image pages a policy block may come from the prompt rather than a face. Keep leading with faces, the likely cause there, but add a short "or check your prompt" so a blocked prompt doesn't send someone to swap out a face-free image. Addresses BILL's TRIVIAL copy finding.
b75c276
benjcooley
left a comment
There was a problem hiding this comment.
👨💻 BILL: APPROVED at b75c2760fc, a scoped re-review of 5039ba2590..b75c2760fc. 0 BLOCKER / 0 SHOULD FIX / 0 TRIVIAL.
The only change is the workshop.error.policyRealFaces copy in translations.ts: en now ends "…or check your prompt.", and zh-CN gets the matching clause. That resolves my trivial. No test hard-codes the sentence (they go through the key), and the rest of the PR is unchanged from my approval at 5039ba2590.
| '模型提供商因内容政策阻止了输入或生成的输出。请检查提示词和参考文件后再运行。' | ||
| }, | ||
| 'workshop.error.policyRealFaces': { | ||
| en: 'Seedance blocks realistic human faces, even AI-generated ones. Try an image without one or a stylized character, or check your prompt.', |
There was a problem hiding this comment.
TRIVIAL (copy, non-blocking): thanks — this resolves my earlier finding. One nit introduced by the edit itself: the comma after "without one" was dropped, and "without X or Y" greedily parses as "without (X or Y)".
Try an image without one or a stylized character, or check your prompt.
reads as "try an image that has neither a realistic face nor a stylized character" — which inverts the recommendation, since a stylized character is the thing we want them to try. The old wording ("Try an image without one, or a stylized character.") was unambiguous; re-adding the clause is what cost it.
The zh-CN line does not have this problem, because it repeats the verb (换一张不含人脸的图片 或使用 风格化角色), so the two options stay separate.
Suggestion — restore the split, e.g.:
Seedance blocks realistic human faces, even AI-generated ones. Try a stylized character, or an image with no face at all. The block can also come from your prompt.
Not blocking; approving either way.
benjcooley
left a comment
There was a problem hiding this comment.
👨💻 BILL: APPROVED at b75c2760fc, a scoped re-review of 5039ba2590..b75c2760fc (not the whole PR).
| Severity | Count |
|---|---|
| BLOCKER | 0 |
| SHOULD FIX | 0 |
| TRIVIAL | 1 |
Scope of the new work: one commit, one file, 2 lines — the workshop.error.policyRealFaces en and zh-CN strings in translations.ts. No logic, no test, no config change. REAL_FACE_REFUSALS, refusesRealFaces, the runSlug threading through ReelEvent → reduceReel, and the policyMessage prop on PlaygroundOutput are all byte-identical to what I approved at 5039ba2590.
Prior finding resolved. My TRIVIAL was that this copy replaces the generic policy text for every policy block on those seven pages, including prompt-triggered ones, and could send someone with a blocked prompt off to swap a face-free image. Both locales now carry a prompt clause. Thread resolved.
Test impact checked, not assumed. CinematicTakeNotice.test.ts:41 asserts on rendered English via queryByText(/realistic human face/) — it is coupled to this string, contrary to what I implied last round. The new copy still opens with "realistic human faces", so the regex still matches and the test still passes. PlaygroundOutput.test.ts passes its own literal through the policyMessage prop and is unaffected. Flagging the coupling so the next copy edit to this key does not surprise anyone.
One new trivial, inline: re-adding the clause cost the disambiguating comma — "Try an image without one or a stylized character" parses as "without (a face or a stylized character)", inverting the stylized-character suggestion. zh-CN avoids this by repeating the verb. Non-blocking.
CI: green except pr-risk / Grade PR risk, still pending — that is the only reason the PR reads UNSTABLE.
Process note, not a code finding: commit b75c2760 is authored by me (the reviewer), not by @wealthwiselabs. I wrote the fix for my own comment and am now approving it, so this round carries no independent reviewer on those 2 lines. Fine at this size for a copy tweak; worth a second pair of eyes if anything non-trivial lands this way.
ELI5
Seedance's video models won't accept a picture with a realistic human face — even an AI-generated one — but when they refuse, our server only passes back "blocked by content policy," so people are left guessing what to change. Now a blocked Seedance take tells them the actual rule and what to try instead.
Motivation
In Cinematic Studio, a Seedance 2.5 video with a starting frame of a photoreal (AI-generated) person comes back "Blocked by the model's content policy" with generic advice to review the prompt and reference files. BytePlus documents that Seedance 2.5 and the 2.0 series do not accept input images or videos containing real human faces (Create portrait videos with Dreamina Seedance models). The Router forwards this only as
content_policy_violation("The request was refused by the model's content policy"), dropping the provider's face code, so the site cannot say which input caused it. Without this, users cannot tell the block is about the face rather than their prompt.Provenance
400error_type: content_policy_violationwith the generic detail above. A different photoreal AI portrait went through (5 s clip), so the provider's classifier is per image.pnpm vitest runonPlaygroundOutput,ModelDetail,cinematic-studio(components + lib),workshop-model-restrictionsandsrc/composables: 69 files, 808 passed;pnpm typecheck:website: 0 errors;eslinton the changed files andoxfmt --check apps/website/src: cleanReviewer context
InputImageSensitiveContentDetected.PrivacyInformation) through, which would let the site show this only when the face was the causeSummary
workshop-model-restrictions.ts: records the realistic-face refusal for the seven Seedance 2.x pages that take images or video (2.5 edit / first-last-frame / reference, 2.0 image-to-video / reference, 2.0 Fast first-last-frame / reference), with its BytePlus source and review date, beside the existing frame-ratio rule.refusesRealFaces(slug)reads it. Text-to-video pages, Seedance 1.x and other providers are not covered.workshop.error.policyRealFaces(en + zh-CN): "Seedance blocks realistic human faces, even AI-generated ones. Try an image without one, or a stylized character." It states the documented rule without claiming this block was caused by a face, since the Router does not say.runSlug, the operation that ran, because a starting frame or references move a shot off the model it was directed on (Seedance 2.5 → its first/last-frame operation).CinematicTakeNoticeshows the face copy on a policy block whoserunSlughas the rule.PlaygroundOutputtakes an optionalpolicyMessage(same pattern ascancelledMessage);ModelDetailpasses the face copy on the covered pages.Test plan
/models/byteplus--seedance-2-5-reference--generate-videos/, a blocked run shows the same copy; on a Seedance text-to-video page it does not