I kept running into the same wall trying to put two or more characters into one scene — no matter what I tried, their LoRAs would bleed into each other. Faces would drift, features would mix, identities would blur together. So I built this from the ground up — separate passes, zero bleed, by design.
What does your model do?
Generates two distinct AI characters side by side — each with their own LoRA — plus the background, in three fully separate, independent passes. No shared latent between any of them at any point, so there's no identity bleed or feature mixing between the characters' identities and LoRAs. The results are then automatically combined side by side into one shared scene: correctly scaled relative to each other, positioned believably, seams cleaned up, with an optional upscale pass. (Same-scene placement, not physical interaction — see Limitations.) All fully automated across 130 nodes and 27 organized groups — no manual cropping, positioning, or compositing required.
What's it for?
Producing clean, multi-character still images — especially as source frames for downstream image-to-video generation (e.g. Wan 2.2), where getting the composition, scale, and identity fidelity right before animating matters far more than fixing it after, since I2V models faithfully preserve whatever they're given, flaws included, and Wan2.2 immediately fixes images, adding detail, lighting, etc afterward.
What is your model good at?
Zero identity bleed — each character is generated in total isolation from the other, so likeness stays exactly what each LoRA intends
Auto-scaling & positioning — measures each generated figure and matches proportions/placement automatically; no manual cropping or dragging
Fully configurable per-character controls — independently adjustable for each person:
Height — how tall/close they appear, as a % of canvas
Ground line — vertical foot placement
Center offset — horizontal position measured from canvas center (not a fixed edge), so resizing never silently shifts someone off-target
Resolution-independent — every size/position value is percentage-based, driven from one control, so it holds up at any output resolution
Configurable seam quality — optional edge feathering and seam-only color matching (deliberately scoped to avoid recoloring the whole subject)
Full control over outputs — every render gives you a guaranteed unmodified raw version, plus a processed version reflecting whichever refinement stages you've enabled (edge cleanup, seam color/lighting match), and optional 2x and 4x upscales on top. Toggle each stage independently — nothing is locked into a single fixed output.
Model flexibility — swap between a standard checkpoint and an alternate (SFW/NSFW) checkpoint from one toggle; optional extra LoRA slots per character
Requirements at a Glance
Base model: Krea 2 Turbo (or Lustify/Krea 2 v1.0, or any nsfw checkpoing, for the optional NSFW mode)
Custom node packs: ComfyUI Essentials, KJNodes, rgthree-comfy
Manual downloads: SAM3 checkpoint, RealESRGAN upscale model
Full install instructions, direct download links, and troubleshooting steps for a known Desktop install issue are all included inside the workflow itself (see the "Required Nodes, Models & Troubleshooting" note).
What is your resource bad at? / Limitations
No physical interaction between characters. Because each person is generated in a completely separate pass with no awareness of the other, this workflow places them side-by-side in a shared scene — it cannot produce contact, touching, hugging, or any pose where the two characters physically interact. This is the direct tradeoff for zero LoRA bleed: the same isolation that keeps identities from mixing also means neither pass can "see" or react to the other's pose. Perhaps in a future update this can be solved. But not an easy problem to fix. Due to this, the best use case is probably for generating images to toss into i2v video models since then dual character interaction can happen with perfect identify preservation.
Fine hair-strand detail is limited by the segmentation mask's precision — hair reads as a clean silhouette rather than individual strands
No lighting-direction matching between the two characters and the background beyond seam-level color correction — scenes with strong directional light (e.g. sunset) may show subtle inconsistency
This is a genuinely complex, many-node workflow — read the in-graph "🚀 START HERE" note before diving in
NSFW checkpoints + side-profile male characters: SAM3's segmentation can struggle with side-profile poses on NSFW-oriented checkpoints like Lustify, producing less reliable masks. Front-facing poses work consistently. A fix is being investigated for a future version. Female characters have no issues like this.