RS MultiCharacter Sampler (Hangeul)

RS MultiCharacter Sampler

Progressive resolution: RS Progressive Resolution Sampler samples larger images faster by starting at 0.5–1.0 of the final side lengths and expanding with Lanczos. This trades some detail for speed; lower start scales can lose more detail or change composition, and quality is not guaranteed. Actual speed depends on the model and settings. Existing multi-character sampling remains available.

Regional generation and local detailing for ComfyUI. Give each character a prompt, an optional LoRA, and a point, box or mask to guide where their attributes appear.

한국어 · 日本語 · Documentation · Changelog · Releases · 简体中文 · 繁體中文

Current release: 1.1.3 Hangeul. Runs with local models; no RS NovelAI extension or API key is required.

Open the app screen — RS Workspace (experimental)

Included with 1.1.0 and later. No separate app installation is needed.

1. Install or update RS MultiCharacter Sampler in ComfyUI-Manager. Check that the installed package is 1.1.0 or later. Git users can run git pull --ff-only inside this extension's folder.
2. Restart the ComfyUI server, then reopen its usual browser page and press Ctrl+Shift+R (macOS: Cmd+Shift+R).
3. Click RS Workspace at the bottom right of ComfyUI. You can also use Edit → RS Model Workspace (Experimental) where the frontend exposes extension commands.
4. Choose an installed model on the left. The app selects that model's loaders, starting settings and supported feature routes. Enter character prompts, adjust placement, then click 생성 (Generate).
5. Click 노드 그래프 (Node graph) to return to the editable workflow. Opening the app from an ordinary workflow creates a separate tab and keeps your existing work.

The app currently uses Korean labels; the English guide explains them. See the complete Korean walkthrough for wildcards, masks, 3D poses, camera, references, LoRAs, face detailing and saving. Missing files or unsupported feature combinations are shown before generation. Base models are installed separately.

Prompt autocomplete: a lightweight offline fallback is included. Enabled Custom Scripts AutoCompleter, Autocomplete-Plus and Autocomplete++ take priority in the original node textareas, including the app. See autocomplete details.

No button? Check the installed version, restart the server as well as refreshing the browser, and confirm the extension is enabled in ComfyUI's extension settings. Keep a single installation. If a reverse proxy is used, load ComfyUI through its normal working URL. Troubleshooting.

1.0.10: Trained Krea 2 Turbo and MOSS 3D pose adapters work alongside numbered character regions. auto prefers native JSON for active MOSS pose control, and regional routing for Krea. First-use weights download automatically. Setup and limits · 36-image audit.

Camera control: RS Camera Control adds smoothly blended camera captions to shared RS/NAI contexts, with an interactive view editor. It supports random views and optional region zoom, rotation and flat-plane perspective; generated 3D pose remains model-dependent.

Camera reference: Anima camera-tag helper / ComfyUI-KR-Camera-Control (credits ComfyUI_bsk_UI / 灰暗x). Reference scope and notices.

Unified 3D pose control (experimental): connect 3D Pose Masks → building pipe → Finalize → Sampler. Anima 2B, Krea2 and MOSS load their matching adapter internally. No separate MODEL/VAE/patch node is needed. Select masks_only to disable learned guidance; unsupported architectures use masks only. Exact joints/contact are not guaranteed. Setup and limits.

What you can do

• Write a shared scene and separate character prompts, with up to 10 character/text regions.
• Assign character LoRAs and edit placement with points, boxes, external masks or a numbered mask-painting canvas.
• Turn colored OpenPose PNG, JSON or POSE_KEYPOINT data into editable person masks.
• Detect and locally detail characters using their original prompts and matching LoRAs.
• Save PNG/WebP generation settings and restore them through the image loader.

Coordinates and masks guide prompt regions. Exact people counts, identity, poses and text spelling remain model-dependent. Character image references are also available as an experimental, not recommended option; prefer a matching character LoRA when available.

Anima Role K/V (experimental): the Image Reference node can optionally prepare independent K/V banks for numbered character regions. Its dedicated 504 MiB adapter is downloaded automatically on first use. Default reference-frame behavior is retained. Setup, routing limits and validation.

Anima 2B / 2.9B reference frames: select either checkpoint; the sampler automatically prepares its matching reference adapter. Model choice remains yours. The separate Role K/V adapter currently targets 2B only. Setup and limits.

Anima reference v2 · experimental: explicitly select Auto: Anima v2 (2B) on the MODEL adapter node to reuse reference K/V throughout denoising. It is faster in the measured cases but can lose hair/clothing identity and use more memory. It does not replace the default pilot path; v2 2.9B and character LoRAs are unsupported. Setup and comparisons.

Install or update

Find RS MultiCharacter Sampler in ComfyUI-Manager, or use the Comfy Registry listing. Package ID: comfyui-rs-multicharacter-sampler.

For Git installation, run this inside ComfyUI/custom_nodes:

git clone https://github.com/raspie10032/ComfyUI_RS_MultiCharacter_Sampler.git

Update through Manager, or run git pull --ff-only inside an existing Git installation. Restart ComfyUI and refresh the browser after installing or updating. Keep one installation of this extension.

Python 3.11+, Torch, NumPy and Pillow are required. Your ComfyUI core must natively support the selected model family. Supply the diffusion model, matching text encoder and VAE yourself. Missing krea2 or moss_image choices in Load CLIP require a core/environment update. See compatibility and tested versions.

Optional features have separate setup:

| Feature | Setup | | --- | --- | | Local detailer | Install ComfyUI-Impact-Pack and ComfyUI-Impact-Subpack through Manager, then install requirements-detailer.txt with ComfyUI's Python. SAM is optional. Selected missing detector/tagger catalog models download when executed. | | Image references · experimental | Anima reference frames, Krea 2 Identity Edit v2 and native FLUX references (FLUX.1 needs Kontext); SD1.5/SDXL/MOSS keep their existing paths. Install requirements-reference.txt. Active references download weights; OFF/strength 0 loads none. Backends and limits |

Start with a workflow

Load the modular two-character Anima workflow, or choose your model in the table below. Use the UI JSON; files ending in -api.json are for API clients. The original workflow PNG can also be dropped onto the canvas.

1. Select the installed MODEL, CLIP and VAE in the loader nodes. For the Anima example, keep Load CLIP type stable_diffusion.
2. Connect STRING text nodes to RS Character Prompt, one character's appearance and action per numbered row. Write the scene, style and total person count in the sampler's common prompt; use its common negative prompt for exclusions.
3. Set RS Character Position. X/Y range from 0 to 1, with (0, 0) at the upper left. Match the same row numbers in Prompt, LoRA and Position.
4. Connect the building nodes to RS Finalize Character Pipe, then its character_pipe output to RS MultiCharacter Sampler. LoRA and Position are optional; the role nodes can be connected in any order before Finalize.
5. Choose Model family preset → Apply preset, check resolution and model-specific settings, then run. Presets copy sampling/regional values; choose models, LoRAs, prompts, layout, dimensions and seed separately.

Autosave writes to output/RS_MultiCharacter/ as PNG, lossless WebP or lossy WebP. It stores generation information; use ComfyUI's ordinary Save Image if you also want the editing workflow embedded.

Character building guide · Preset values · Sampler/scheduler comparisons for Anima

Models and examples

RS NAI interoperability: share character/context pipes, detection combine loaders and taggers with the updated RS NAI extension. Prompts and positions transfer; local masks/LoRAs and each backend's settings survive round trips. Behavior and limits.

These are example starting points, not universal quality settings. FLUX variants have separate workflows and sampling requirements.

| Model family | Workflow | Starting point and scope | | --- | --- | --- | | Anima | Two characters | Main example; matching encoder and Qwen Image VAE. Default sampler pair: Euler ancestral / SGM uniform. | | SDXL / Illustrious | Two characters · Character LoRA | Load Checkpoint supplies MODEL/CLIP/VAE. Basic example: 30 steps, CFG 5.5, Euler ancestral / normal. Tested derivatives and limits | | FLUX.1 / 2 / 3 | Model-specific workflows | Includes Dev, Schnell and Klein variants. Variant matrix and validation | | Krea 2 Turbo | Two characters · Character LoRA | 8 steps, CFG 1, Euler / simple; Qwen3-VL 4B (krea2) + Qwen Image VAE. Common negative has no effect at CFG 1. | | MOSS-Image SFT | Base · 8-step LoRA | Base: 50 steps, CFG 3, Euler / simple; Qwen3.5 9B (moss_image) + Flux2 VAE. 8-step setup requires a separate acceleration LoRA. | | SD1.5 · experimental, unstable | Two characters | Execution tested; people count and attribute separation are unstable. Trial and limits |

Krea 2 and MOSS regional support is experimental. MOSS uses ordinary ComfyUI CFG, without the official pipeline's negative-branch layer-22 skipping. Assets and validation

Character LoRAs and adapters

Choose each character's file and model/CLIP strengths in RS Character LoRA. A standard ComfyUI LoRA loader before the sampler applies a global LoRA. The checkpoint, adapter architecture and dimensions must match.

• Precision is the default: independent predictions for each LoRA setting, with factor-only optimization where supported.
• Normal shares layer calculations for supported Anima/SDXL linear LoRAs. Unsupported models/adapters automatically use Precision and report the reason. The modes can produce different character details.
• Supported static formats include LoRA, LoCon/Tucker, DoRA, LoHa/DoHa, LoKr/DoKr, GLoRA, OFT/BOFT, IA³ and matching diff/set patches. Packed INT8/FP8 paths include partial Q/K/V LoRA on MOSS. Complex adapters can cost more time and temporary memory.

GGUF: connect the MODEL output of the separately installed ComfyUI-GGUF loader. Regional LoRAs preserve packed weights; FLUX uses Precision, including automatic fallback from Normal. See setup and verified scope.

Native quantization: packed FP8/INT8 regional LoRAs use the connected loader’s native operations across supported model families. Kernel availability and image quality depend on the quantized release, ComfyUI and GPU. See formats, tests and limits.

T-LoRA, GLoKr, other custom tensor types and shape-changing patches are outside the adapter path's supported scope. Model/CLIP restrictions and numerical checks are documented in the adapter format guide. See also LoRA setup, mode comparisons and tested character-LoRA combinations.

Placement, masks and local detailing

Use RS Character Position to drag numbered points, boxes and masks. Use RS Character Mask for external MASK inputs, or RS Character Mask Drawing to paint over a display-only reference image. Mask guide and examples

For two people plus a word, use three slots: two character point regions and one text box region. Connect the literal word to the text row's STRING input and leave its LoRA at None. For an object, use a character box slot with an object description. Use layout_method=regional; MOSS native layout requires boxes for every region. Mixed-layout examples · Object examples · Layout guide

RS OpenPose Masks (Experimental) builds approximate person masks; it does not send a skeleton into the model. Optional body-part name guidance defaults OFF and supports Anima/SDXL. RS Character Interaction can delay separation decay outside overlapping regions on Anima/SDXL and experimental SD1.5. Exact joints and contact are not guaranteed. Pose inputs, previews and limits · Pose example

For local detailing, load the generation + detailer workflow. Connect the image, character pipe and same MODEL/CLIP/VAE to RS MultiCharacter Detailer, plus RS Detection Combine Loader and RS Character Tagger Loader. Start with face SEGM, WD ViT v3 and denoise 0.25. Tagging selects an original character prompt; it does not append inferred tags. Check the reported character assignments. Execution checks and limits

Restore generation settings

Open an original saved PNG/WebP in RS MultiCharacter Load Image and connect its character pipe to the sampler. Image prompts and characters replace the common prompt fields. Enable use_image_settings to also restore seed, dimensions and sampling/regional controls; disable it when applying a preset or using current widgets.

Model, LoRA, CLIP and VAE connections must be supplied explicitly. Metadata does not download or load those assets. Reference images must be reconnected. This is text-and-seed regeneration; the detailer separately performs masked img2img. NAI V4-style character captions can be imported, but their sampling settings cannot be replayed locally. Metadata and restoration boundaries

Anima reference-frame experiment: Connection and limits. This separate Anima-only adapter supports one global reference or independent numbered character references, while retaining regional prompts. It requires separately downloaded weights. Character reference separation is experimental and costs additional predictions.

Node reference

The extension registers 15 nodes. UI labels and contextual help follow ComfyUI's English, Korean, Japanese, Simplified Chinese or Traditional Chinese locale.

| Nodes | Purpose | | --- | --- | | RS Character Prompt / LoRA / Position / Finalize Character Pipe | Build numbered character settings and the final sampler/detailer pipe. | | RS Character Mask / Mask Drawing | Connect external masks or paint numbered regions. | | RS 3D Pose Masks (Experimental) | Drop a single-person PNG/JPEG/WebP onto its Image → 3D panel: estimate, save the scene and open the 3D editor automatically. This replaces the node's scene. Import inside a group editor to replace only its selected person. A one-person scene uses the image's camera and crop; off-frame body parts are excluded. First use downloads SAM 3D Body (2.83 GB, separate SAM License); requires ComfyUI core SAM3DBody support. | | RS OpenPose Masks (Experimental) / Character Interaction | Build pose-derived masks and adjust overlap separation. | | RS Image Reference (Experimental · Not Recommended) | Associate an image reference with a character number. | | RS MultiCharacter Sampler / Load Image | Generate and restore saved settings. | | RS MultiCharacter Detailer / Detection Combine Loader / Character Tagger Loader | Detect, match and locally correct characters. | | RS MultiCharacter Prompts · legacy | Combined input retained for existing workflows. Use the modular nodes for new workflows. |

Limits and troubleshooting

• One image per execution, up to 10 input slots. Dense layouts and crossing props can duplicate or omit people, mix attributes or form panels. A single character owns the full canvas; moving its only point does not move the person.
• The regional sampler uses a common negative prompt. Per-character negative captions and scheduled text conditioning are unsupported by the sampler.
• SD2, inpainting/refiner checkpoints, FLUX Fill/Control and external ControlNet/attention patches are unsupported by the regional sampler. The optional image-reference feature has its own documented backends and limits.
• The sampler accepts 2–100 steps and CFG 1–20. Presets do not identify the connected model or guarantee quality. Model-specific workflows are the best starting point.
• Restoration requires the original image metadata and matching assets/runtime. Social platforms can strip metadata. Autosave does not retain unrelated EXIF/GPS/camera data.
• For errors, include the extension/core versions, exact checkpoint and LoRA versions, workflow JSON and full traceback in an issue.

Documentation and validation index separates generation trials, numerical adapter tests and CPU/CI compatibility checks.

Development and license

With Torch, NumPy and Pillow installed, run from the repository root:

python -m unittest discover -s tests -p 'test_*.py' -v
node tests/test_ui.cjs

Set RS_MC_TEST_COMFYUI to a ComfyUI checkout for native integrations. The CI workflow lists all twelve JavaScript UI suites and the core/OS matrix.

GPL-3.0-only. See COPYING, third-party notices and source lineage.