RS MultiCharacter Sampler (Giyeok Fix)

ComfyUI_RS_MultiCharacter_Sampler

0.2.1 Giyeok Fix (기역 픽스) · Anima · SDXL · Krea 2 · MOSS · SD1.5 (experimental, unstable) — independent prompts with X/Y, box and mixed regions for local ComfyUI generation.

Giyeok Fix: fixes character and text inputs appearing together. Each slot now uses one shared STRING socket, preserving its connection when kind changes. Old workflows move their literal-text connection to that socket. Fix and compatibility.

New in Giyeok: separate Character Prompt, Character LoRA, Character Position and Finalize Character Pipe nodes. Each role shows all active rows together; connect the three roles in any order. Positive prompts use STRING connections, with no individual UC input. Position and Finalize show layout guides. Changes and validation scope · Usage.

Each slot uses one STRING: a character prompt or the literal words to draw, selected by kind. Changing kind keeps the same connection.

Zeta Fix: fixes Anima character LoRAs silently missing the actual inference model with ComfyUI dynamic VRAM. Update and restart ComfyUI; existing LoRA selections and strengths can be reused. Fix, controlled image comparisons and validation scope.

New in Zeta: choose a LoRA and model/CLIP strengths for each character. The sampler blends masked predictions through native ComfyUI hooks; the detailer uses the matched character's LoRA. Usage, cost and limits.

Ordinary Anima LoRAs automatically share small factor matrices through native Bypass LoRA; other supported paths retain weight hooks. This saves expanded-weight storage, while distinct LoRA combinations still require separate predictions.

0.1.6 release notes and validation scope.

New in Epsilon: mixed point/box layouts, literal text regions, eight editable model-family presets, MOSS native layout mode, and two-person object-box examples for Anima and Krea 2. See release notes and validation scope.

한국어 · 日本語 · Architecture · Validation

This extension contains six nodes. It does not require the RS NovelAI extension, an API key, or a ControlNet model. It uses your supported local model and matching text encoder.

| Node | Purpose | | --- | --- | | RS MultiCharacter Prompts | 1–10 regions with original text, enable switch, normalized X/Y and optional boxes or literal text. One dedicated RS_MC_CHARACTER_PIPE output. | | RS MultiCharacter Sampler | Common positive/negative prompts, local sampling, optional autosave, image/latent/pipe/info outputs. | | RS MultiCharacter Load Image | PNG/WebP image and mask, restored character pipe and generation information. | | RS MultiCharacter Detailer | Detect regions, match original character prompts and locally correct masked crops. | | RS Detection Combine Loader | SEGM / BBOX / SAM selection in one detector connection; SEGM takes priority. | | RS Character Tagger Loader | Local tagger selection for automatic character matching, with download progress. |

Epsilon layouts: optional boxes and mixed point/box layouts, text regions, and MOSS native layout comparison. Setup and limitations.

Mixed-layout examples now include Anima, Illustrious XL, Krea 2 Turbo and experimental SD1.5, with model-specific conditioning and prompts. Anima's initial example has been retuned for a shared campus background and an unclipped title. Settings, comparisons and limits. Text spelling and person count remain model-dependent; these settings do not automatically replace existing workflow values.

The SD1.5 example checks two distinguishable people and visible lettering in the requested area. Natural image quality and exact spelling are outside that compatibility criterion.

To try two people plus one word, use region count 3, layout_mode=mixed, point geometry for character slots 1–2, and kind=text with literal HELLO for slot 3. Set its X/Y and width/height, then select sampler layout_method=regional. Runnable examples · Validation and visual limitations. Up to 10 total character/text regions; MOSS native mode requires boxes for every region.

Two people plus one object: use the existing Anima — mixed or Krea 2 Turbo preset with the 20 object workflows. Slot 3 uses character + box with an object description. The ten-object trial preserved two people and recognizable objects in all 20 images, but exact bounds and scene coherence varied. Object slots are not automatically excluded from face detailing. Images, settings and limits.

Modular character input

Use separate all-prompts, all-LoRAs, and all-positions nodes. Each role shows rows 1–N together, with matching row numbers identifying characters. Connect these roles in any order, then finish with Finalize Character Pipe for Sampler or Detailer. Positive prompts use external STRING connections with no character UC field. Position and Finalize show a layout guide. Usage in English / 한국어 / 日本語 · Two-character example. The old all-in-one node remains deprecated for saved workflows.

Install

Find RS MultiCharacter Sampler in ComfyUI-Manager, or open the Comfy Registry page. Package ID: comfyui-rs-multicharacter-sampler. Restart ComfyUI after installation. If the new listing is not yet visible in your Manager cache, refresh its node list or use the Git installation below.

For Git installation, clone this repository into your ComfyUI custom_nodes directory, then restart ComfyUI. Use one installation method; do not install a second copy alongside an existing Git checkout.

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

Run the command inside ComfyUI/custom_nodes. The sampler uses ComfyUI's existing PyTorch, NumPy and Pillow. Python 3.11+ and a ComfyUI version supporting your chosen model are required; the tested environment is recorded in the validation report. Supply your own diffusion model, text encoder and VAE. The optional detector/tagger loaders download their selected missing catalog models when executed.

For detailing, install ComfyUI-Impact-Pack and ComfyUI-Impact-Subpack through Manager, then install requirements-detailer.txt using your ComfyUI Python environment. Restart ComfyUI. SAM is optional and must already be installed to select it. Additional tagger choices can require the optional dependencies listed in their error message.

SD1.5 — experimental, unstable

Ordinary SD1.5 checkpoints can run through the existing sampler. Character counts and attribute separation are unstable. The base checkpoint completed four two-character requests; two images contained three people, with some hair/clothing attributes mixed or omitted. This is limited execution evidence, not a success-rate estimate or validation of SD1.5 fine-tunes. See the trial settings and scope and workflow.

Experimental SDXL generation

Load the two-character SDXL workflow, then select your checkpoint in Load Checkpoint. Branch its MODEL / CLIP / VAE outputs into the existing sampler. Illustrious XL v1.0 and NoobAI XL epsilon-pred 1.1 were tested with real generation and prompt-position swaps. This does not establish compatibility across all SDXL derivatives or sampler combinations.

The example starts at 30 steps, CFG 5.5, Euler ancestral + normal, initial common prior 0.3, character gain 1.3, spatial separation 0.5. Existing Anima defaults stay unchanged. Its 0.05 / 1.7 / 0.9 settings can cause visible panel boundaries on SDXL.

XY routes prompts, not exact body positions or guaranteed counts. Multiple characters and grid layouts can duplicate people, mix attributes or create panels. SDXL Refiner, inpainting models, ControlNet and other attention patches are unsupported. PNG/WebP metadata identifies the SDXL engine and restores through Load Image. See method and live results.

Experimental Krea 2 / MOSS-Image support

Use a current ComfyUI with native support for both models. If Load CLIP lacks krea2 or moss_image, update ComfyUI and its environment dependencies first. Choosing a model does not automatically change sampling settings.

| Model | Example | Text encoder / VAE | Starting settings | | --- | --- | --- | --- | | Krea 2 Turbo | Two characters | Qwen3-VL 4B, krea2 / Qwen Image VAE | 8 steps · CFG 1 · Euler + simple | | MOSS-Image SFT | Two characters | Qwen3.5 9B, moss_image / Flux2 VAE | 50 steps · CFG 3 · Euler + simple |

Connect MODEL, CLIP, VAE and the existing character pipe. Write the scene in the common prompt and each character's appearance and action in natural-language sentences. At Krea 2 Turbo's CFG 1, the common negative has no effect. The MOSS example uses ordinary ComfyUI CFG; it does not apply the official pipeline's negative-branch layer-22 skipping.

The sampler accepts 2–100 steps; Anima/SDXL retain their 30-step defaults. XY selects prompt regions and does not guarantee exact counts, poses or complex interactions. This support covers T2I. Reference-image editing and detailer quality on these new models have not been validated. See model files, terms and validation scope.

MOSS 8-step LoRA

Use the two-character or four-character XY workflow with LoRA 0.7 · 8 steps · CFG 3 · Euler / simple. Download lora_latest.safetensors from lodestones/MOSS-Image-SFT-8-steps-dmd into ComfyUI/models/loras/, refresh the model list, then select the downloaded filename in Load LoRA (Model Only). Connect Load Diffusion Model → Load LoRA (Model Only) → RS MultiCharacter Sampler; CLIP and VAE keep their normal connections.

Use MOSS-Image SFT, a Qwen3.5 9B encoder (Load CLIP type moss_image), and Flux2 VAE. Updating this extension alone does not add MOSS to an older ComfyUI core. Update the core and its Python dependencies, then restart. ComfyUI 0.41.0 was used for the LoRA trial.

CFG 3 gave clearer character attributes in the two tested scenes; CFG 1 was faster but weakened some colors. Eight steps at CFG 3 used 16 model evaluations; CFG 1 used 8. Neither setting guarantees counts or props. CFG 1 workflows, exact assets, comparisons and limits.

Local multi-character detailing

Load the generation + detailer example. Branch the same MODEL / CLIP / VAE (including your LoRA) to the sampler and RS MultiCharacter Detailer. Connect the sampler's image and character_pipe outputs to the detailer, then connect RS Detection Combine Loader and RS Character Tagger Loader. An existing image can instead supply image and character_pipe from RS MultiCharacter Load Image.

Start with the default face SEGM, WD ViT v3 and denoise 0.25. The basic view contains denoise, seed and autosave; expand the options for crop size, mask settings, matching thresholds and sampling. SEGM skips BBOX and SAM entirely. With SEGM OFF, choose BBOX alone or BBOX + SAM. OFF models are not loaded. Existing models in ultralytics/segm, ultralytics/bbox and sams appear when you refresh the model list.

The tagger only selects an original character prompt; inferred tags never enter the correction prompt. Hole filling and notch closing are enabled by default. Crops use the original image, with separate ownership masks and feathering inside those masks. Ambiguous regions are preserved and the report explains the reason. High-confidence misidentification is still possible; inspect the reported character assignments, especially when several characters look alike. The pipe has up to 10 characters; the detailer processes up to 10 regions by default, 64 maximum. More processed regions do not guarantee quality.

Detailer autosave keeps the original generation settings and adds a separate correction history in PNG/WebP metadata. It does not embed a workflow. PNG/lossless WebP preserve pixels outside the applied mask; lossy WebP can change pixels through compression. Anima has the broader validation set. SDXL (Illustrious XL v1.0), Krea 2 Turbo and MOSS-Image SFT each also completed one two-face detailer run; these are basic execution checks, not quality or success-rate comparisons. See detailer validation.

First workflow

Drop the original workflow PNG onto the ComfyUI canvas, or load examples/two-characters-anima.json. The shared PNG contains ComfyUI workflow and prompt metadata; the sampler's normal autosave behavior is unchanged. The separate -api.json file is for API clients.

1. Select your Anima diffusion model in Load Diffusion Model, its text encoder in Load CLIP, and its VAE in Load VAE. The example uses anima-base-v1.0.safetensors, qwen_3_06b_base.safetensors and qwen_image_vae.safetensors. Select the names actually installed on your machine. Use the example's CLIP type stable_diffusion.
2. Write the overall scene/style and total person count in the sampler's common prompt. Write one character's appearance/action in each character slot; the tested prompts use 1girl or 1boy per slot.
3. Set X/Y from 0 to 1. (0, 0) is the upper left. Equal centers spread in slot order from left to right. Lower character_count excludes hidden slots even if their old text remains saved.
4. Select Model family preset → Apply preset on the sampler to copy the tested starting values for Anima, SDXL/Illustrious, Krea 2 Turbo, MOSS base/8-step LoRA or experimental SD1.5. Anima and SDXL have separate XY and mixed choices. You can edit every value afterward. Models, prompts, layout, dimensions and seed remain your choices. Preset values and usage
5. Run. Autosave writes to output/RS_MultiCharacter/. Choose PNG, lossless WebP or lossy WebP. Standard model LoRAs can be connected before the sampler with ComfyUI's LoRA loader.

The IMAGE output can also go to ComfyUI's ordinary Save Image node. The sampler's own autosave stores generation metadata, not a ComfyUI workflow. Use ordinary Save Image if you need workflow embedding.

Actual output from the bundled API example, without the private style LoRA.

Sampler and scheduler choices

Every sampler and scheduler registered in your ComfyUI installation is selectable. Start with the default pair below, or try the two alternatives. The dropdown tooltips also list these starting points.

| Starting point | Sampler | Scheduler | Observation in this trial | | --- | --- | --- | --- | | Default | euler_ancestral | sgm_uniform | Clear character features and shared background | | Alternative composition | euler | linear_quadratic | Preserved features and followed the requested face-to-face seating | | Clear alternative | dpmpp_2m_sde_gpu | beta | Distinct hair, outfits and cups; seated side by side instead of opposite |

These observations come from Anima Base v1.0, one two-character cafe scene and seed, 1536×768, 30 steps, CFG 4, no LoRA. All 45×9=405 combinations completed generation; several produced severely degraded images. These are starting points, not a ranking or success rate across prompts.

Select both fields yourself. Other combinations remain available, and the node never substitutes a recommended choice. Native option names are preserved for workflows and image metadata. Defaults remain euler_ancestral + sgm_uniform; the available inventory depends on your ComfyUI version.

Some CFG++ variants degraded at the fixed test settings. Lower-CFG/fewer-step retests have not been performed, so this does not establish incompatibility. The sampler accepts 2–100 steps and CFG 1–20. The Anima recommendations still refer to the original 30-step comparison; lower-CFG/fewer-step CFG++ recovery is untested. See matrix results and recommendation evidence.

Restore from an image

Select an original saved image in RS MultiCharacter Load Image and connect its character_pipe to the sampler. Image prompts and characters replace the sampler's common prompt fields. Enable use_image_settings to also restore seed, dimensions, CFG, sampling and region settings; leave it disabled to use the sampler widgets for those values.

Model/LoRA/CLIP/VAE connections remain explicit. File names and strengths are recorded in generation_info and metadata, but the loader does not load or download them. Reproduction requires the same assets and compatible runtime. This path regenerates from text and seed; it is not img2img.

The loader also accepts NAI V4-style character captions with one center per character. Select NovelAI prompt syntax for text typed in that syntax; imported NAI images select it automatically. NAI settings cannot be replayed as local settings: keep use_image_settings off for those images.

Current limits

• Anima, SDXL, Krea 2, MOSS and unstable experimental SD1.5 support, one image per execution. The sampler is text-to-image; the detailer runs masked img2img crops. SD1.5 has only been tested for generation, not detailing. SD2, Flux, inpainting/refiner checkpoints, ControlNet and other attention patches are not supported by the regional sampler.
• 1–10 input slots. Actual image tests cover Anima 1–6, SDXL 2–6, Krea 2 / MOSS 2–4, and SD1.5 two-character requests. More characters do not guarantee quality. Horizontal layouts are the strongest tested use case; successful new-model 2x2 examples do not establish a general success rate.
• Coordinates choose prompt regions; they do not lock a head/body to a point or enforce exactly one person per region. A single character owns the whole image, so moving that one center does not move the person.
• Multiple rows, dense interactions and crossing props can duplicate or omit people. Four requested people in a 2×2 experiment became 12 and 10. Six people in two rows became 7 and 11. Do not treat count or identity as guaranteed.
• Per-character negative captions are preserved in metadata but rejected by the regional sampler in this release. Use its common negative prompt. The detailer can use the matched character's negative caption. Scheduled text conditioning is not supported.
• The regional sampler accepts 2–100 steps, with 30 as the Anima/SDXL default. The detailer has its own adjustable step count. Initial common weight, character gain and boundary release are advanced experimental controls, not universal quality settings.
• Autosave does not preserve an uploaded image's unrelated EXIF/GPS/camera data. Social platforms may strip generation metadata; use the original file for restoration.

See validation for real generation, metadata replay and the limits of the evidence. This release retains the documented count, identity and interaction limits.

Development

From this directory, using an environment containing Torch, NumPy and Pillow:

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

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

ComfyUI compatibility

SDXL does not require Anima, Krea 2 or MOSS to exist in ComfyUI. Each newer architecture still requires a ComfyUI checkout with native support for that model. SD1.5 also passed native CPU integration at all three snapshots below; see its limited real-generation results. The earlier compatibility audit’s CPU API integration tests cover ComfyUI v0.3.76 (SDXL), 15eb748 (Anima/SDXL/Krea 2) and 5cc4b87 (all four families). These are tested snapshots, not a guarantee for every intervening version or third-party model patch. The compatibility report separates real generation, small native model tests and Windows CPU CI.