项目的 README,来自 github.com/ukr8b3g-cmyk/H3-Character-Sheet-Designer。合并为一个 markdown 文件:README.md。
日本語 · Package v1.5.1
Create character reference sheets with native MiniMax H3 image-reference conditioning. One character image and up to three optional part references are passed through unchanged. Select up to 13 views: six head views, three full-body views, front/back clothing, paired hands and paired feet/footwear.
A MiniMax H3 Image GPU test at 1888 × 1792 with a fixed seed produced all six angles in the expected positions, with the upper horn tips visible. The rightmost hair touched the frame in that test. With matching inputs and dimensions, both nodes produce the same Head 6 prompt and guide. The models can generate different results.
In MiniMax H3 Image, the 11-view No clothing · natural body surface layout assigns each portrait to a specific slot when Use Layout Image is on and the full-body side prompt is edited. The prompt requests white gaps between views and keeps the shared reference stance instruction scoped to unedited full-body views. The clothing menu tooltip explains how the mode treats clothing panels, hands, and footwear.
On ComfyUI_W, a GPU run with the same wolf reference, seed 42, and a 4224 × 1280 canvas used Use Layout Image on and an edited standing side-view prompt. The finished sheet had no visible guide lines, all three full-body views showed the wolf standing on two legs, and all 11 requested views appeared exactly once. Results can vary with the reference and seed. Issue #6 · Release notes
No clothing · natural body surface is now available under Clothing shape in MiniMax H3 Image and P-H3 v1.3. Choose it for unclothed animal characters or mascots to keep the body structure, fur, colors, and markings from the reference. Hand and footwear settings still apply independently.
Both workflows produced 11 views, including a bipedal full-body side view and separate hand and foot close-ups, in a GPU test with the same wolf reference, seed 42, and 4224 × 1280 output. View count and pose can vary with the reference and seed. Issue #6 · Release notes
For MiniMax H3 Image, Standard · 5 views in Auto size with Use Layout Image enabled gives each full-body panel about 6% more width and leaves 10.5% above and at least 3.5% below the figure in the guide. The two portrait panels give up a little width to keep the sheet dimensions fixed.
Fixed-seed GPU comparisons at 3168 × 1280 produced all five views for both reference images. The horned character fit from horn tips to shoe soles in all three full-body views. Framing can vary with the reference, seed, and output size. Release notes
A fixed-reference, fixed-seed GPU comparison produced one view of bare hands and one view of feet in black socks, with a white gap. Composition and fine details can vary. The existing workflow files and node names are unchanged, including the P-H3 workflow. Restart ComfyUI after updating.
Requires H3 Character Sheet Designer v1.4.0+ and ComfyUI 0.45.0+ for the new MiniMax H3 Image workflow. Update ComfyUI if MiniMax H3 Image Conditioning is missing, then restart and reload the browser after updating the Designer.
| Workflow | Designer node | Model path |
|---|---|---|
| MiniMax H3 Image Character Sheet Designer | MiniMax H3 Image Character Sheet Designer v1.4.0 | MiniMax H3 Image Conditioning and three image-model files |
| H3 Character Sheet Designer — P-H3 Subgraph | H3 Character Sheet Designer v1.3.0 | P-H3 Image Conditioning and its original model set |
Use the exact three files selected in the template. Set CLIP type to minimax. Select your character image in Load Image before running. The workflow uses the supplied image as a generation reference.
| Component | Download | Folder under ComfyUI/models/ |
|---|---|---|
| Diffusion model | minimax_h3_image_pruned_int8_convrot.safetensors | diffusion_models/minimax/ |
| Text encoder | qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors | text_encoders/ |
| VAE | minimax_h3_image_vae_fp16.safetensors | vae/ |
ComfyUI's MiniMax H3 Image guide
| Component | Download | Folder under ComfyUI/models/ |
|---|---|---|
| Diffusion model | minimax_h3_fl2va_pruned_int8_convrot.safetensors | diffusion_models/minimax/ |
| Text encoder | p-h3-image_text_encoder_int8_convrot.safetensors | text_encoders/ |
| VAE | p-h3-image_vae_int8_convrot.safetensors | vae/ |
| LoRA | p-h3-image_lora_bf16.safetensors | loras/P-H3-Image/ |


13-view GPU example, 4832 × 1280, Layout Image OFF. The requested views appeared once each in this run. Fine details and viewpoint adherence can vary with reference, seed and resolution. 8 MP and 16 MP can require more VRAM and time.
Package v1.3.1 keeps the existing H3 Character Sheet Designer v1.3.0 and v1.2.2 node names. In the v1.3.0 node, open Prompts after Settings, select one of 13 views and edit its H3 base instruction. Apply saves it; Reset this prompt restores that view's default. Blank overrides are allowed. Valid pending edits are committed on workflow save/queue. Each override accepts up to 2,000 UTF-16 units.
These are H3 instructions, not Anima tags. Layout, image references, Part prompts and settings are added separately. This edits a view's base instruction rather than the whole final prompt. Overrides persist in saved workflows. Unedited workflows retain their original instructions. Existing workflow files, node IDs/titles, sockets, image passthrough and Layout Image ON/OFF are unchanged. The previous v1.2.2 node does not gain this editor.
Validation: local implementation passed Python 176 and JavaScript/DOM 112 tests. Live Chrome/ComfyUI_W checks covered Apply, Reset, save/reload and Layout Image ON/OFF; live preview API checks covered reference numbering. This update adds no GPU quality, speed or VRAM benchmark. Human remnants in clothing panels and missing/duplicate views remain possible; custom edits may weaken viewpoint compliance. Restart ComfyUI and reload the browser after updating.
None for the guide image.No human, and flat-outfit instructions scoped to clothing panels. Clothing-only output retains outfit features rather than person identity. Clothing and all-13 results have not demonstrated a major general improvement. Human remnants, missing/duplicate views and outfit/full-body confusion remain possible.See release notes and CHANGELOG. Previous changes remain recorded there.
P-H3 template and archived releases: P-H3 subgraph · v1.3 · v1.2.2. The P-H3 workflow uses minimax_h3_fl2va_pruned_int8_convrot.safetensors with the P-H3 image text encoder and VAE.
The template requires native ComfyUI H3/subgraph support and rgthree-comfy for Power LoRA Loader. The Designer adds no Python runtime dependency. Pika Ref2VA is enabled at 1.0, FL2VA is disabled, and ManualSigmas has eight intervals. These are the supplied comparison settings, not a universal optimum or the original Pika six-NFE sampling contract. Keep the supplied save format unless you intentionally change it.
Set the CLIP loader type to minimax and select an H3 encoder. Music3 encoders are incompatible and cause sampler matrix-shape errors (for example 32768 versus 5120). Model switching/first loading is separate from layout compilation.
| Component | Download / filename | Folder under models/ |
|---|---|---|
| Diffusion / template | DaSiWa Hybrid V4 INT8 — DasiwaMinimaxH3_dasiwaHybridV4_int8.safetensors |
diffusion_models/minimax/ |
| Native alternative | H3 Ref2VA INT8 — minimax_h3_ref2va_pruned_int8_convrot.safetensors |
diffusion_models/ |
| Text encoder / CLIP | H3 Qwen3-VL INT8 — qwen3vl_32b_minimax_h3_int8_convrot.safetensors |
text_encoders/ |
| Alternative encoder | H3 NVFP4 AWQ — qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors |
text_encoders/ |
| Video VAE | H3 VAE INT8 — minimax_h3_video_vae_int8_convrot.safetensors |
vae/ |
| Alternative VAE | H3 VAE FP16 — minimax_h3_video_vae_fp16.safetensors |
vae/ |
| Pika LoRA / enabled | Ref2VA — pika_h3_flex_ref2va_v1_bf16.safetensors |
loras/minimax/ |
| Pika LoRA / disabled | FL2VA — pika_h3_flex_fl2va_v1_bf16.safetensors |
loras/minimax/ |
Choose the adapter matching your base model and conditioning path. Both Pika links are provided; this template enables Ref2VA only. Follow model licenses/access requirements. The minimax/ subfolder is template organization, not a required ComfyUI convention.
| Designer output | Native H3 input |
|---|---|
prompt, width, height |
matching prompt/dimension inputs |
character_image |
ref_images.ref_image_0 |
layout_image |
ref_images.ref_image_1 |
reference_image_1 / 2 / 3 |
ref_images.ref_image_2 / 3 / 4 |
Keep the guide wire connected when OFF. Picture numbering skips None images. The four image pass-through outputs retain the input tensors unchanged. The template samples five frames and saves the first decoded frame as the static sheet.
In Part prompts assign extra references to Same character, Face, Head/face/hair, Complete outfit, Outfit front/back, Hands or Footwear. Explicit choices and part instructions take priority, then dedicated references, then the baseline image; later dedicated references of the same role take priority. These are text-based permissions, not regional attention isolation. Separate panel descriptions and No human do not guarantee that instructions stay inside one panel.
| Node ID / display version | Purpose |
|---|---|
H3CharacterSheetDesignerReferenceV13 / v1.5.1 |
Current P-H3 workflow |
H3CharacterSheetDesignerReference13 / v1.2.2 |
Previous behavior for existing Reference13 workflows / comparison |
These are the previous node versions. Earlier Designer, Reference, Multi Reference and Reference V1.1 IDs are not registered in v1.3. Their historical source/workflow files and old releases remain available; use an earlier release if those IDs are required. Saved workflows and user node titles are not automatically rewritten. V1.2.2 preserves its compiler/state behavior; its shared UI also has the Worn switch removed.
Download the v1.5.1 package. It includes the Designer and both workflow JSONs.
H3-Character-Sheet-Designer directly under ComfyUI/custom_nodes. The __init__.py file should be directly inside the Designer folder.workflows folder. Install the model files listed in that workflow's Model Links note and select them in the model loaders. The MiniMax H3 Image workflow requires ComfyUI 0.45.0 or later.For a Git installation, run these commands from the directory that contains ComfyUI:
cd ComfyUI/custom_nodes
git clone https://github.com/ukr8b3g-cmyk/H3-Character-Sheet-Designer.git
ComfyUI/custom_nodes as a backup. Extract the v1.5.1 package to the original location so ComfyUI loads one copy of the Designer. For an existing Git clone on main, use git pull --ff-only in that clone.workflows. Check that the Designer node title ends in v1.5.1 and that Head · 6 views shows three columns and two rows.v1.5.0 changes the Head · 6 views layout, guide, and prompt, so an existing workflow can open with a different composition. For the original behavior, use an older package with the workflow from the same release.
ComfyUI/custom_nodes as a backup. Extract the archive and put the folder containing __init__.py directly in ComfyUI/custom_nodes. Keep one installed copy of the Designer there.workflows folder. The MiniMax H3 Image template requires ComfyUI 0.45.0 or later and the image-model files in the folders listed above.MiniMax H3 Image版とP-H3版の Head · 6 views は、Layout ImageのON/OFFとも3列×2段です。プレビュー、プロンプト、配置ガイドで各ビューの位置と画面上の向きを揃えました。ONでは枠線のないガイドを使い、上段に余白を取ります。同じ入力条件・寸法なら両版のHead 6のガイドとプロンプトは一致し、生成結果はモデルによって変わります。
衣服なし/体表を保持では、白い余白で区切った淡い6面で配置を示します。プレビューとガイドは同じ配置を使い、生成時の顔の向きは各面のプロンプトで指定します。
配布ワークフローは MiniMax H3 Image と P-H3 Subgraph の2本です。現行ノードの表示は両方とも v1.5.1 です。
ComfyUI/custom_nodes/H3-Character-Sheet-Designer/__init__.py となるように配置してください。workflows から使うJSONを開きます。ワークフロー内のModel Linksにあるモデルを導入し、各ローダーで選択します。MiniMax H3 Image版にはComfyUI 0.45.0以上が必要です。custom_nodes の外へバックアップし、同じ場所にv1.5.1の本体を展開します。Gitで導入している場合は、そのクローンの main で git pull --ff-only を実行できます。v1.5.0ではHead · 6 viewsの配置・ガイド・プロンプトが変わるため、既存のワークフローを開いても、旧版と同じ構図や生成結果になるとは限りません。旧版と同じ挙動が必要な場合は、同じリリースのパッケージとワークフローを組み合わせて使ってください。
ComfyUI/custom_nodes の外へ移動します。ZIPを展開し、__init__.py を含むフォルダーを ComfyUI/custom_nodes の直下へ置きます。導入するDesignerは1つにします。workflows フォルダー内にあるJSONを開きます。MiniMax H3 Image版にはComfyUI 0.45.0以上と、上記のモデルファイルが必要です。MiniMax H3 Image版の 衣服なし/体表を保持 で、Layout Image ON かつ全身側面のプロンプトを編集した場合に、顔6面の位置を個別に指定し、ビュー間は白い余白で分けるよう指示します。参照と同じ姿勢を求める共通指示は、未編集の全身ビューにだけ適用します。衣装メニューのツールチップにも、このモードの衣装ビュー・手袋・履物の扱いを追記しました。
通常のComfyUI_Wで、同じオオカミ参照画像・seed 42・4224 × 1280・Layout Image ON を使用し、側面の編集文で二足の立ち姿を指定してGPU検証しました。ガイド線は出ず、正面・側面・背面の全身3面はすべて二足の立ち姿で、手足のアップを含む11面が各1枚揃いました。結果は参照画像やseedによって変わります。 Issue #6 · リリースノート
MiniMax H3 Image版とP-H3 v1.3版の 衣装の型 に 衣服なし/体表を保持 を追加しました。服を着ていない動物やマスコットの体の構造、毛並み、色、模様を参照画像に合わせます。手袋と履物はそれぞれの設定で指定できます。
同じオオカミ参照画像・seed 42・4224 × 1280でGPU検証し、両版で全身側面が二足の姿勢を保ち、手足のアップを含む11面が揃いました。ビュー数や姿勢の追従は参照画像やseedにより変わります。 Issue #6 · リリースノート
MiniMax H3 Image版で Standard · 5 views、Autoサイズ、Use Layout Image ON を選ぶと、ガイドの全身3枠を各約6%広げ、上に10.5%、下に3.5%以上の余白を取ります。顔アップ2枠から幅を振り分けるため、シートの出力寸法は変わりません。
3168 × 1280のGPU比較では、2枚の参照画像で5ビューが各1枚ずつ揃い、角のある人物も全身3面で角先から靴底まで収まりました。収まり方は参照画像、seed、出力サイズによって変わります。 リリースノート
同じ参照画像・seedでのGPU比較では、素手と黒い靴下の足が白い余白を挟んで各1面出ました。配置や細部は生成条件により変わります。P-H3版を含む既存ワークフローとノード名は従来どおりです。更新後はComfyUIを再起動してください。
新しいMiniMax H3 Image版には、この拡張のv1.4.0以上と、ComfyUI 0.45.0以上が必要です。 MiniMax H3 Image Conditioning が見つからない場合はComfyUIを更新してください。
minimax。Load Imageで参照画像を選び、Euler / Simple、6ステップ、CFG 1で生成します。標準のSave Animated WEBPを1 FPSに設定し、静止画1枚を保存します。パッケージは1.3.1ですが、ノード名 H3 Character Sheet Designer v1.3.0/v1.2.2 は保持します。v1.3.0ノードのSettingsの次に Prompts(プロンプト) タブを追加。13ビューから選んでH3基本文を編集し、Applyで適用、Reset this promptで対象だけ初期化できます。空欄も保存可能。各欄2,000 UTF-16単位まで。有効な未適用編集は保存・キュー送信時にも確定します。
AnimaタグではなくH3用の文章です。配置・画像参照・部位プロンプト・設定は別途追加され、最終文全体の置換ではありません。編集はワークフロー内に保存し再読込で復元。未編集の旧ワークフローの生成指示、既存ファイル、ノードID・表示名・入出力・画像受け渡し・Layout Image ON/OFFは維持。旧v1.2.2には編集タブを追加しません。
ローカル実装でPython176件・JS/DOM112件成功。実Chrome/ComfyUI_Wで適用・初期化・保存/再読込・Layout Image ON/OFF、実APIで参照番号を確認しました。今回GPU画質・速度・VRAMの追加評価はしていません。衣装への人物混入やビュー欠落・重複の制約は残り、編集文によって方向の追従が弱まる場合もあります。更新後はComfyUI再起動とブラウザー再読込が必要です。
パッケージv1.3.1(ノード名はv1.3.0/v1.2.2を維持)。MiniMax H3本来の画像参照でキャラクターシートを生成します。人物画像1枚+部位参照3枚をそのまま受け渡し、必要なら配置画像を追加参照にします。MOSS版の画像→文章化とは異なる方式です。
現行P-H3テンプレートと旧リリース — P-H3サブグラフ版 · v1.3版 · v1.2.2版。P-H3版は minimax_h3_fl2va_pruned_int8_convrot.safetensors とP-H3 Image用テキストエンコーダー・VAEを使用します。
配線は上の表と共通です。人物→ref_image_0、配置画像→1、追加参照3枚→2/3/4。NoneはPicture番号から除外。画像Tensorは変更せず返します。5フレームを生成し、最初の復号フレームを静止画として保存します。
テンプレートはrgthree-comfyが必要です。Pika Ref2VAは強度1.0で有効、FL2VAは無効、ManualSigmasは8区間。今回の比較設定であり万能な最適値やPika本来の6-NFE契約を保証するものではありません。保存形式は必要に応じて利用者が変更してください。
モデル/TE/VAE/LoRAのファイル名・URLは上の Model downloads for the older Pika v1.3 template 表にまとめています。Pika Ref2VA/FL2VA両方のリンクを掲載し、このテンプレートはRef2VAだけを有効にしています。CLIPローダーはtype=minimax、H3用Qwen3-VLを使用。Music3用TEは非互換で行列サイズ不一致になります。初回ロードの時間とDesignerの配置計算は別の処理です。各モデルのライセンス・アクセス条件に従ってください。
追加参照の役割は同一人物/顔/頭部・髪/衣装全体/衣装前/衣装後/手/履物。明示設定・部位指示、専用参照、基準画像の順に適用し、同じ専用役割では後の接続を優先します。これはテキスト上の指示であり、領域attentionを完全分離する機能ではありません。
登録は今回版v1.3.0と前版v1.2.2の2つです。前版Reference13の生成動作・状態契約は保持しますが、共有UIのWorn切替は撤去しています。それより前のDesigner/Reference/Multi Reference/V1.1のIDが必要なら旧リリースを使用してください。旧ファイル・Releaseは履歴として残し、保存済みワークフローやユーザーのノード名は自動変更しません。
衣装・13面は小修正で、人物除去・向き・枚数・同一性の大幅改善を実証した版ではありません。写真・アニメ・CG、別seedでは失敗が残ります。ガイドの幾何改善もGPU画質向上の証明ではありません。Auto幅の拡大は時間・VRAM負荷を増やす場合があります。追加LLM、TE文章生成、BREAKによる分離、領域別エンコードは導入していません。
CPU/UI/配線の確認とGPU評価を区別しています。ユーザーの動作確認はありますが、最終変更の複数参照・複数seedでの成功率、VRAM、速度の比較評価は未完了です。今回の公開にはRegistry/Manager登録を含めません。
Keep LICENSE, NOTICE.md and artwork attribution. Model licenses apply separately.