[feature] ComfyUI-on-Modal manga character pipeline for visual novel assets

[lf1up]

Summary

New standalone imagegen/ root directory: a ComfyUI deployment on modal.com that generates consistent 2D manga characters for the visual-novel features — one full-body master sprite per character plus identity-preserving derived shots of the same character (face close-ups with different expressions, upper-body, hands, profile…), "interactive manga adventure" style.

Approach (hybrid, per plan)

• Stage A — base sprite: Animagine XL 4.0 Opt (SDXL anime, CreativeML Open RAIL++-M — commercial use permitted) renders a full-body master sprite; BiRefNet (MIT) cuts a transparent VN-ready PNG.
• Stage B — body parts / expressions: Qwen-Image-Edit-2509 fp8 (Apache 2.0, native ComfyUI nodes) takes the master sprite + an instruction ("Close-up portrait of this exact character's face, smiling…") and renders identity-preserving shots. Lightning 4-step LoRA by default (~4–8× faster), full 20-step schedule via --no-lightning.

What's included

• imagegen/modal_app/app.py — Modal app: pinned image build (comfy-cli 1.12.0 / ComfyUI 0.27.0 / comfyui_birefnet_ll@1.1.4), ~38 GB of HF weights cached in a Modal Volume and symlinked into ComfyUI, interactive ComfyUI dev server (modal serve), headless JSON API (proxy-auth protected) on an L40S with scale-to-zero, and a modal run smoke test.
• imagegen/workflows/ — API-format graphs: character_base.json (txt2img + alpha branch) and character_edit.json (Qwen edit + alpha branch).
• imagegen/scripts/generate_character.py — CLI turning a character spec (flags or JSON file) into a full bundle: base.png, base_raw.png, face_<expression>.png, shot_<name>.png, manifest.json with seeds for reproducibility. 8 built-in expressions, 4 built-in shots, custom instructions supported.
• imagegen/RESEARCH.md — research findings: Modal ComfyUI pattern, model licensing comparison, consistency approaches considered (crop-and-refine, VNCCS, per-character LoRA), future work (persona wiring, poses, memory snapshots).
• imagegen/README.md — setup, deploy, usage, costs (~$0.05–0.15 per bundle on L40S).

Validation

Ran against a live local ComfyUI 0.27.0 installed with the exact pinned commands from the Modal image (CPU flavor):

• comfy --skip-prompt install --fast-deps --version 0.27.0 and comfy node install --fast-deps comfyui_birefnet_ll@1.1.4 install cleanly with comfy-cli 1.12.0.
• All 4 workflow variants (base ± alpha branch, edit ± Lightning/alpha) submitted to the live /prompt endpoint: graph/schema validation passes; the only reported errors are the missing model files themselves (expected locally — they live in the Modal Volume).
• The infer() serving path exercised end-to-end with real comfy run --wait executions: input-image write → execution → output collection by unique prefix → RGBA correctness proven (with InvertMask the foreground stays opaque; the counterfactual without it produces a fully transparent foreground, confirming the JoinImageWithAlpha inversion semantics).
• All 6 HF model files resolve anonymously (no token needed), ~38.3 GB total, verified via HF metadata.
• Modal SDK 1.5.2 API usage verified (fastapi_endpoint(requires_proxy_auth=...), modal.experimental.stop_fetching_inputs, App.cls(scaledown_window=...)); app imports and registers ui + ComfyUI.
• Client CLI helper logic tested (spec merge, expression/shot resolution, argparse).

Not yet run: live modal deploy + GPU sample bundle — blocked on Modal credentials (no ~/.modal.toml / MODAL_TOKEN_ID in the agent environment). Once credentials are available (Cursor Dashboard → Cloud Agents → Secrets, or modal setup locally):

cd imagegen && uv sync
uv run modal deploy modal_app/app.py   # builds image + downloads weights (one-time)
uv run modal run modal_app/app.py      # smoke test: base + happy face to /tmp/imagegen-smoke

Out of scope (documented in RESEARCH.md)

Wiring into agent/ persona data, generating the 9 existing personas, per-character LoRA training, pose ControlNet.

  

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
careersim-landing	Ready	Preview, Comment	Jul 17, 2026 2:14am
careersim-web	Ready	Preview, Comment	Jul 17, 2026 2:14am

[coderabbitai]

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: 908e0760-86b5-4983-bf5b-1b1984262398

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch cursor/comfyui-manga-pipeline-3752

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 72.09%. Comparing base (cf657bc) to head (f560c10).
⚠️ Report is 2 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main      #67   +/-   ##
=======================================
  Coverage   72.09%   72.09%           
=======================================
  Files          60       60           
  Lines        4254     4254           
  Branches      662      662           
=======================================
  Hits         3067     3067           
  Misses        978      978           
  Partials      209      209           

Flag	Coverage Δ
agent	68.59% <ø> (ø)
api	82.77% <ø> (ø)

Flags with carried forward coverage won't be shown. Click here to find out more.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.