CivArchive
    H3 Look Sheets - v1.3 - Relay VSR/SeedVR2
    Preview 149762050
    Preview 149779562

    # H3 Look Sheets v1.3 — character + outfit reference sheets with MiniMax H3

    Put a person photo and a few outfit photos in. Get a contact sheet of that person, wearing that exact look, from every angle.

    H3 Look Sheets is a ComfyUI custom node pack and workflow built on MiniMax H3's multi-reference mode (ref2va). It describes your reference photos with a vision model, writes the shot-by-shot H3 prompt for you, renders one take that moves through every angle you asked for, picks the best frames, and lays them out as a clean look sheet.

    Useful for character consistency, fashion/lookbook sheets, LoRA dataset prep, and costume references.

    📺 Demo videos: https://www.youtube.com/playlist?list=PLR0q7a2fnl5w

    ---

    ## ✨ What's new in v1.3

    - New RelayVSR x2 workflow: same pipeline, with a RelayVSR x2 upscale (core ComfyUI nodes) instead of SeedVR2. It upscales the whole video right after H3 Face Refine, before Select Frames. Needs ComfyUI v0.45.0 or later.

    - The SeedVR2 workflow is renamed ComfyUI-H3LookSheets-2StageSamplingUpscaleSeedVR2.json.

    ## ✨ What's new in v1.2

    - Shot Config is simpler: one shot_description combo (8 body regions × 8 camera orientations, e.g. "Lower Legs - Back") replaces the separate angle/framing combos. Expression is only written when the face is in frame. Experimental.

    - Cleaner takes: the prompt asks for an instant hard cut to a static, locked-off camera on every shot and holds the last shot until the end. It also states that an outfit photo's own camera angle and pose aren't kept, only the garment.

    - Per-shot frame selection: Select Frames gets a per_shot scope that picks frames inside each shot, driven by the new shot_count output of the prompt node.

    - New nodes: Prompt View (colorized prompt preview, copy button, editable override), Ref Image Size (match/max combo that survives a Set/Get pair), Datasheet Images (choose which images go into the sheet, from the flow and/or added by hand).

    - Post-processing in the example workflow: H3 Face Refine, upscale and datasheet. A "Model Links" note lists every model and custom node used.

    - New standalone Datasheet workflow to build a sheet from images you already have.

    - Fix: stale CUDA graph replay crash in Describe Reference.

    ## v1.1 recap

    - Up to 8 outfit photos (front/back of top and bottom, shoes, up to 4 accessories), each described only for its own piece.

    - Several person photos as identity references; hair stays identical across shots.

    - Image Aggregator routes everything into H3's 9 reference slots.

    - Automatic gender detection, plus retries/fallback on refused or empty descriptions.

    ---

    ## 🔧 How it works

    1. Load the person photo(s) and the outfit photo(s).

    2. Describe Reference: a Qwen3-VL vision model writes one sentence per image (person: face, hair, body; outfit: garments and accessories only).

    3. Shot Config ×1–15: pick the shot description (body region + camera orientation) and expression for each shot.

    4. Look Sheet Prompt - Custom Shots writes the full H3 ref2va prompt, with every <Picture N> tag matching the wired images. Check or tweak it in Prompt View.

    5. MiniMax H3 renders the take (2-stage sampling + latent upscale).

    6. Optional post-processing: H3 Face Refine, then SeedVR2 or RelayVSR x2 upscale.

    7. Select Frames picks the keepers, over the whole take or per shot.

    8. Datasheet Images lets you switch frames on/off or add your own; Datasheet Settings lays them out as one contact sheet.

    ## 🧩 Nodes

    | Node | Role |

    |---|---|

    | Look Sheet Prompt - Custom Shots (H3) | Builds the ref2va prompt from 1–15 shots, outputs shot_count |

    | Shot Config (H3 Look Sheet) | Shot description / expression for one shot |

    | Describe Reference (H3 Look Sheet) | Vision description of up to 8 photos, with retry + fallback |

    | Image Aggregator (H3 Look Sheet) | Routes person + outfit images into H3's 9 reference slots |

    | Prompt View (H3 Look Sheet) | Colorized prompt preview, copy, override |

    | Ref Image Size (match/max) | ref_image_size picker that survives Set/Get |

    | Select Frames (H3 Look Sheet) | Picks the best, most varied frames, whole take or per shot |

    | Datasheet Images (H3 Look Sheet) | Chooses which images go into the sheet |

    | Datasheet Settings (H3 Look Sheet) | Lays frames out as a contact sheet |

    ---

    ## 📦 Requirements

    - MiniMax H3 ref2va (MiniMaxH3ReferenceToVideo, core ComfyUI node). The example workflows use the int8 ref2va model + the 4-step turbo LoRA.

    - Qwen3-VL as the vision CLIP (the example uses qwen3vl_4b_fp8_scaled) for Describe Reference and, optionally, Select Frames tier 2.

    - RelayVSR workflow: ComfyUI v0.45.0 or later (RelayVSR nodes are core since then; no custom node to install). The SeedVR2 workflow runs on v0.44.0.

    - Post-processing uses extra models and custom nodes (latent upscaler, H3 Face Refine, SeedVR2 or RelayVSR, KJNodes, Impact Pack…): all listed with links in the workflow's "Model Links" note.

    - No extra Python dependencies.

    ## 💾 Install

    cd ComfyUI/custom_nodes

    git clone https://github.com/shisa84/ComfyUI-H3LookSheets

    Already installed? Update with:

    cd ComfyUI/custom_nodes/ComfyUI-H3LookSheets

    git pull

    Restart ComfyUI and refresh the browser. The nodes appear under H3LookSheets. Drag a workflow from example_workflows/ onto the canvas (reload it after updating: a graph already saved keeps its old copy).

    | Workflow | What it is |

    |---|---|

    | ...2StageSamplingUpscaleSeedVR2.json | Full pipeline with SeedVR2 upscale |

    | ...2StageSamplingUpscaleRelayVSR.json | Full pipeline with RelayVSR x2 upscale (ComfyUI v0.45.0+) |

    | ...Datasheet.json | Datasheet only, for images you already have |

    ## 💡 Tips

    - Keep your references consistent: same person, same hair across all person photos. Inconsistent inputs give inconsistent sheets.

    - Crop the head off outfit photos worn by a different model, so their face or hair doesn't leak into your character.

    - Assign every outfit image to a slot (top / bottom / shoes / accessory_N). An unassigned close-up (e.g. earrings) gets described as "nothing visible".

    - Using per_shot in Select Frames? Wire shot_count from the prompt node: a wrong count merges or splits shots.

    - Low VRAM? Favor medium and close-up shots over wide ones. The 2-stage workflow renders small first, then upscales.

    ---

    🔗 GitHub: https://github.com/shisa84/ComfyUI-H3LookSheets/tree/main

    Description

    • The example workflow is renamed ComfyUI-H3LookSheets-2StageSamplingUpscaleSeedVR2.json.

    • New ComfyUI-H3LookSheets-2StageSamplingUpscaleRelayVSR.json: the same pipeline with a RelayVSR x2 upscale (core RelayVSR nodes) in place of SeedVR2. It upscales the whole video right after H3 Face Refine (saved as a video), before Select Frames. Needs ComfyUI v0.45.0 or later — see Requirements.

    FAQ

    Workflows
    MiniMax H3

    Details

    Downloads
    103
    Platform
    CivitAI
    Platform Status
    Available
    Created
    11/21/2026
    Updated
    11/22/2026
    Deleted
    -

    Files

    ComfyUI-H3LookSheets-2StageSamplingUpscaleRelayVSR.json

    ComfyUI-H3LookSheets-2StageSamplingUpscaleSeedVR2.json