comfyui-workflow-recipes

name	comfyui-workflow-recipes
description	Curated comfyui workflow recipes Helios fires via the hydra-creative.comfyui.* MCP tool. Each recipe lists input schema, output, when-to-use, and cost.
user-invocable	false
argument-hint	<list|pick|run>
allowed-tools	["Read"]

ComfyUI Workflow Recipes

A curated registry of canonical comfyui workflows Helios can dispatch. Each recipe is a stable JSON workflow with a documented input schema, predictable output, and a rough cost estimate.

Purpose

Keep workflow knowledge in one place so Helios and its sub-crew can pick the correct recipe for an AssetJob without re-inventing graphs per shot. Recipes are versioned and live as procedural resources (low-risk class, auto-evolve).

When to use

• Helios is converting a Shot from a ShotList into an AssetJob and needs to pick a workflow.
• A new visual style requires a workflow that does not exist yet — propose addition.
• A render fails and a different recipe is being evaluated.

Inputs

• The Shot entry (subject, mood, camera angle, lens, lighting notes)
• Destination channel and aspect ratio
• Color-space target (from color-science)

Outputs

• Selected recipe id + filled input schema
• Cost estimate for the render
• Predicted output format (PNG/JPG/MP4/MOV), resolution, color space

How a recipe is invoked

Via hydra-creative.comfyui.run (existing comfyui agent reused; declared as a tool dependency in squad.yaml):

hydra-creative.comfyui.run(
  recipe_id="key-art-portrait",
  inputs={...recipe input schema...},
  output_dir="RLM/output/photo/{campaign_id}/",
  c2pa_sign=true
)

Helios never calls comfyui directly except through this MCP path; the sub-crew is delegated by Helios only. All renders pass through governance-c2pa for IP and brand-safety checks before publish.

Canonical workflows

1. key-art-portrait

• When to use — single-subject hero stills (talent, founder, customer); editorial portraits; key art for campaigns where a face anchors the frame.
• Input schema:
  • subject_prompt (string) — describe the subject; do NOT reference a real person's likeness without brand-safety clearance.
  • style_prompt (string) — populated from color-science mood entry.
  • negative_prompt (string).
  • aspect_ratio ("1:1"|"4:5"|"9:16"|"3:4"|"2:3"|"16:9").
  • resolution (int, long edge px; default 2048).
  • seed (int|null).
  • face_fix (bool, default true).
• Output — PNG, sRGB, embedded ICC, C2PA manifest attached.
• Cost estimate — ~$0.05-0.15 per image (model-dependent); face-fix pass adds ~30% time.

2. product-hero

• When to use — product-on-seamless or product-in-context stills for e-com, paid ads, PR kits.
• Input schema:
  • product_reference_image (path) — required; clean cutout preferred.
  • scene_prompt (string) — environment, surface, lighting.
  • style_prompt (string).
  • aspect_ratio, resolution, seed.
  • controlnet ("canny"|"depth"|"none", default "canny" for shape fidelity).
  • shadow_pass (bool, default true) — generate a separate shadow plate for compositing flexibility.
• Output — PNG main + optional shadow plate; sRGB.
• Cost estimate — ~$0.10-0.25 per shot; controlnet adds ~20%.

3. cinematic-keyframe

• When to use — narrative-driven still frames intended to be extended into motion via video-synth sub-agent; mood boards; pitch decks.
• Input schema:
  • scene_prompt (string) — composition, subject, action.
  • style_prompt (string) — anamorphic, film stock emulation, etc.
  • camera_angle (enum from shot-list-protocol vocab).
  • focal_length_mm (int).
  • aspect_ratio (default "2.39:1" cinematic).
  • resolution (default 2560 long edge).
  • seed.
  • lut_family (string|null) — from color-science mood entries.
• Output — PNG, Rec.709 or P3 (per lut_family), C2PA-signed.
• Cost estimate — ~$0.10-0.20 per frame.

4. style-transfer-loop

• When to use — apply a brand-defined style to a batch of stock or existing assets; refresh older creative without re-shoot.
• Input schema:
  • source_images (list of paths).
  • style_reference (path) — brand-style reference image; MUST be brand-owned or licensed (run brand-safety first).
  • strength (float 0.0-1.0, default 0.6) — higher = more style, less source fidelity.
  • preserve_subject (bool, default true).
  • output_format ("png"|"jpg").
• Output — one transformed image per source; same dimensions.
• Cost estimate — ~$0.04-0.08 per image; batches discount via shared model load.

5. ip-adapter-character-consistency

• When to use — generate multiple shots of the same fictional character (mascot, illustrated talent, internal avatar) with visual consistency across a series.
• Input schema:
  • character_reference_images (list of 3-8 paths) — MUST be brand-owned IP or licensed; governance-c2pa rejects copyrighted-character references.
  • scene_prompts (list of strings) — one per output shot.
  • style_prompt (string).
  • aspect_ratio, resolution, seed_base (int).
  • consistency_weight (float 0.0-1.0, default 0.75).
• Output — N images, one per scene_prompt, with consistent character identity.
• Cost estimate — ~$0.08-0.18 per output image; reference encoding amortized per batch.

Picking a recipe

Decision rules:

Shot kind	Recipe
Human portrait, single subject	key-art-portrait
Product on white or in context	product-hero
Narrative frame, intended for motion	cinematic-keyframe
Refresh batch in brand style	style-transfer-loop
Repeated character across series	ip-adapter-character-consistency

If no recipe fits, escalate to Helios; do not improvise an unversioned workflow inside a campaign.

Procedure

1. Read the Shot and identify the shot kind.
2. Pick recipe per decision rules.
3. Fill the recipe input schema. Populate style_prompt, negative_prompt, aspect_ratio, and color-space target from color-science and shot-list-protocol.
4. Validate references: any *_reference_image and any style_reference MUST pass brand-safety IP-clearance before submission.
5. Estimate cost; if > $200 cumulative for the campaign so far, trigger media-cost-cap HITL gate per squad.yaml.
6. Dispatch via hydra-creative.comfyui.run; await completion.
7. Route output through governance-c2pa for C2PA signing and final brand-safety check.
8. Write outputs to RLM/output/photo/{campaign_id}/; log episode to memory with tags=["render","comfyui",recipe_id].

Failure modes / escalation

• Recipe absent for a shot kind — propose new recipe; HITL approves; once approved, register as procedural resource (low-risk, auto-evolve).
• Reference image fails brand-safety — strip and regenerate; if no clean reference available, abandon shot or escalate.
• Render cost exceeds gate — pause; HITL via media-cost-cap.
• Output color space drifts from spec — re-export with explicit color-management node; re-run color-science validation.
• Repeated face artifacts — re-run with face_fix true and different seed; if persistent, switch base model.

References

• Color-management: color-science
• Shot grammar: shot-list-protocol
• Governance: brand-safety, governance-c2pa sub-agent
• Memory tags: render, comfyui, recipe-id