ComfyUI-Krea2-MultiLoRA-Composer Introduction
ComfyUI-Krea2-MultiLoRA-Composer helps you place multiple LoRA-based characters in one shared Krea 2 image. You can assign up to five character LoRAs to editable areas of a canvas, describe each character separately, and write one scene prompt for everything they share, such as the setting, lighting, and camera angle.
A LoRA is an add-on that guides a model toward a particular character or visual style. When you apply several character LoRAs to an entire image, their traits can mix: faces may become hybrids, one character may inherit another’s features, or a subject may disappear. The Composer addresses this by directing each character LoRA mainly to the area where that character belongs.
Two independently routed character LoRAs sharing a balcony scene
This example is an AI-generated demonstration using independently obtained LoRAs. No LoRA weights are included with the extension. The depicted characters belong to their respective rights holders; the project is unaffiliated with and is not endorsed by those rights holders.
How ComfyUI-Krea2-MultiLoRA-Composer Works
Think of the canvas as a shared stage. The scene prompt describes the stage and the action taking place on it. Each character row describes one performer, and its colored box marks where that performer should appear. The Composer combines these instructions into one positive prompt and prepares a single Krea 2 generation.
Three routing methods work together:
- LoRA routing confines each character LoRA’s direct image contribution to its assigned region.
- Subject-token isolation links a character’s LoRA contribution to that character’s words in the prompt, reducing confusion between subjects.
- Regional attention guidance encourages each subject description to influence its intended part of the image.
For example, you could place a red-coated character on the left of a balcony and a blue-coated character on the right. Describe the balcony, evening light, and camera framing once in the scene prompt. Put each character’s trigger, coat, and pose in that character’s own row, then position the two boxes over their intended locations.
The image is still created in one diffusion pass, not assembled from separately generated character cutouts. The base Krea 2 model handles the overall scene, so the characters can share lighting, props, and space. Routing reduces identity mixing, but it cannot guarantee perfect separation: overlapping boxes, small faces, LoRA quality, prompts, and the random seed still affect the result.
ComfyUI-Krea2-MultiLoRA-Composer Features
Character editor and placement
The Composer provides one editor for one to five character LoRAs. Choose the character count, select Set + reset, and use the searchable selector to choose a Krea 2 LoRA for each row.
For each character, you can set:
- Trigger and description: Use the LoRA’s exact training trigger, then describe that character’s appearance, clothing, pose, and interaction with shared objects.
- Strength: Start at
1.0. Adjust it if the LoRA’s influence needs tuning; the best value depends on the LoRA. - Denoising schedule: Start with
0.0–1.0to let the LoRA contribute throughout generation. You can change the interval to control when its influence is active. - Placement region: Drag the colored box to move it and its lower-right handle to resize it. Start with separated boxes that are only large enough to cover the intended characters.
- Region controls: Feathering softens region edges;
0.08is a starting value. When boxes overlap, thenearestpolicy assigns overlapping image positions according to the closest region center.
The placement canvas follows your chosen output aspect ratio, helping you position characters in the same proportions as the final image.
Shared scene prompt
Use Scene prompt for instructions that apply across the image: framing, background, common objects, lighting, atmosphere, and relationships between characters. Keep identity-specific details in the corresponding character rows. This separation makes it easier to change one character without rewriting the whole scene.
The Composer exposes its combined positive prompt as an output. For Krea 2 Turbo, phrase your desired result positively; do not put “avoid” or other negative-prompt instructions into the scene or character fields.
Optional Canvas LoRA
Canvas LoRA is disabled by default. It gives you a separate LoRA for the environment or an overall rendering style, with its own searchable selection, trigger, description, strength, and denoising schedule. Unlike a character LoRA, it has no draggable box.
Choose one of two coverage modes:
| Coverage | What it does | When to try it |
|---|---|---|
| Unboxed area only | Applies the Canvas LoRA outside the character regions, fading beneath their softened edges. | You want a background style without directly applying it to the boxed characters. |
| Entire canvas | Applies the Canvas LoRA across the image while character LoRAs remain assigned to their regions. | You want a consistent style across both scenery and characters. |
Keep character boxes reasonably tight in Unboxed area only mode: scenery inside a box is excluded from that Canvas LoRA. A Canvas LoRA strongly trained on a character may still introduce that subject or its traits into the background. Its trigger, training, strength, and the seed remain important.
SuperSampler and output size
The Krea2 Multi-LoRA SuperSampler samples the Composer’s latent image, decodes it with the Krea 2 VAE, and returns an image at your requested canvas size. It also provides the higher-resolution working image, the high-resolution latent, and diagnostics. Its controls are familiar if you have used ComfyUI’s KSampler.
The Composer’s supersample_scale increases the internal working resolution without changing the requested final dimensions:
| Scale | Approximate internal pixel cost | Practical use |
|---|---|---|
1.0 | 1.00× | Faster tests and ordinary generation |
1.25 | 1.56× | First step to try for more working facial detail |
1.5 | 2.25× | More working resolution at greater memory cost |
2.0 | 4.00× | Selective use when you have enough VRAM |
For example, a 1216 × 832 canvas at 1.25 is sampled internally at 1520 × 1040 and returned at 1216 × 832. Supersampling can provide more working detail, but it does not guarantee a better likeness and uses more VRAM.
For Krea 2 Turbo, the author suggests this starting point in the SuperSampler: 8 steps, CFG 1.0, euler sampler, simple scheduler, denoise 1.0, and lanczos downscaling. Downscaling can also use bicubic, area, or bilinear filtering.
Scene sharing
The Composer can export a portable scene JSON containing canvas dimensions, the supersampling plan, prompts, LoRA settings, placement coordinates, schedules, routing controls, and validation warnings. Use Import JSON for a saved file or Paste JSON for clipboard text. The older krea2_character_router_share_v1 scene format remains supported.
Basic node connections
To render and save an image, connect the nodes as follows:
- Connect your Krea 2 model and matching CLIP to the Composer.
- Connect the Composer’s
model,conditioning,latent, andsupersample_planoutputs to the SuperSampler’smodel,positive,latent_image, andsupersample_planinputs, respectively. - Connect the matching Krea 2 VAE to the SuperSampler’s
vaeinput. - Connect empty conditioning to the SuperSampler’s
negativeinput. Krea 2 Turbo does not use negative prompting, but this input remains part of the sampler interface. - Connect the SuperSampler’s
imageoutput to Preview Image or Save Image.
Use the Multi-LoRA SuperSampler when you want automatic resizing to your requested output dimensions. A standard KSampler connected directly to the Composer’s latent instead returns the larger working resolution when supersampling is enabled.
ComfyUI-Krea2-MultiLoRA-Composer Models
The extension works with Krea 2 Raw or Krea 2 Turbo; these are base-model options, not models supplied by the extension. Use the matching Krea 2 CLIP and VAE, along with character LoRAs trained for Krea 2. Choose the base model that matches the workflow and LoRAs you intend to use. The documented Turbo starting settings appear above; the source documentation does not specify separate recommended sampler settings for Raw.
No character LoRA weights are included in this project. Canvas LoRA is an optional role for a LoRA you select, not an additional bundled model.
What’s New with ComfyUI-Krea2-MultiLoRA-Composer
- Version
0.4.0: The visible project and node names changed to the Multi-LoRA Composer naming. The internal ComfyUI identifiers,Krea2CharacterRouterandKrea2SupersampledKSampler, stayed the same so existing workflows could continue to load. - Version
0.5.0: An optional Canvas LoRA configuration was added. Older workflows load with Canvas LoRA disabled, and olderkrea2_character_router_share_v1scene files can still be imported without modification.
The provided documentation does not include a fuller version-by-version changelog.
Troubleshooting ComfyUI-Krea2-MultiLoRA-Composer
Characters look blended or inherit each other’s traits
- Check that each row uses the correct Krea 2 LoRA and its exact training trigger.
- Keep each character’s appearance and clothing details in that character’s own row.
- Move the boxes apart where possible and make them only large enough to cover their intended subjects.
- Start with strength
1.0, schedule0.0–1.0, feather0.08, and overlap policynearest. - Try another seed or LoRA pair. Routing reduces collisions, but LoRA training and the base model still affect identity.
A character is missing or appears in the wrong place
- Confirm the intended character row is enabled and has the correct LoRA and trigger.
- Check that its colored region covers the location and size you want in the final composition.
- Describe its position and relationship to other subjects clearly in the shared scene prompt.
- Test with less region overlap and a different seed.
The Canvas LoRA does not affect part of the background
Check its coverage mode. Unboxed area only deliberately excludes areas inside character boxes, even when those areas contain scenery behind a character. Tighten the boxes, or try Entire canvas if the style should extend through those areas.
The background contains features from the Canvas LoRA’s character
A LoRA trained primarily on one subject may still encourage that subject to appear. Check its trigger and description, reduce or adjust its strength, and try a different seed. If the goal is an environment style, the LoRA’s training content also matters.
The image is larger than the width and height you selected
Check which node produces the image you are viewing. At a supersample_scale above 1.0, the Composer’s latent is intentionally larger. Connect it to the Krea2 Multi-LoRA SuperSampler with the supersample_plan input to receive an image automatically resized to the requested dimensions.
Generation uses too much VRAM
Lower supersample_scale toward 1.0, or reduce the target canvas dimensions. The approximate internal pixel cost reaches 2.25× at scale 1.5 and 4.00× at scale 2.0.
Frequently asked questions
Can I use more than five characters?
The documented editor supports one to five character LoRAs.
Will supersampling always improve faces?
No. It increases working resolution, but likeness also depends on the LoRA, prompt, seed, pose, and face size.
Can routing distinguish a person from a face on a poster behind them?
Not reliably when both occupy the same two-dimensional region. A face inside a character box can inherit that region’s identity.
Can I reopen an older shared scene?
Yes. The documented krea2_character_router_share_v1 files remain supported, including files from before the optional Canvas LoRA was added.
Why is there a negative input when using Turbo?
It keeps the SuperSampler compatible with ComfyUI’s KSampler-style connections. Connect empty conditioning; Krea 2 Turbo does not use negative prompting.
Learn More about ComfyUI-Krea2-MultiLoRA-Composer
- Krea 2 by Krea AI provides background on the underlying model.
- FreeFuse informed the project’s token-separation and attention-routing direction. The Composer is an independent ComfyUI implementation designed around Krea 2.
- The ComfyUI-Krea2-MultiLoRA-Composer repository is the project source for its documentation and examples.
The extension is licensed under MIT.
