Kroma v0.1 — LoRA for Krea 2

Kroma is a LoRA fine-tune for Krea 2, packaged as a single ComfyUI-compatible safetensors file. It was rank reduced down to rank 256 / alpha 256 and also carries the fully-fine-tuned normalization and modulation tensors as weight deltas, so it reproduces the original fine-tuned behaviour when loaded on top of the base model.

• File: kroma-v0.1.safetensors (~1.88 GB)
• Type: LoRA (264 adapters) + .diff weight deltas (159 RMSNorm / modulation tensors)
• License: MIT

Note on the format: unlike a "pure" LoRA, this file also includes full-weight deltas for the trained RMSNorm scales and modulation tensors (stored as .diff). ComfyUI's LoraLoader applies these automatically, so no special handling is needed.

---

Running in ComfyUI

1. Requirements

• A recent ComfyUI with native Krea 2 support (update to the latest master).
• The Krea 2 diffusion model, text encoder, and VAE.

2. Files to download

Place these in your ComfyUI folders:

What	Where	Notes
kroma-v0.1.safetensors	ComfyUI/models/loras/	this repo
Krea 2 diffusion model	ComfyUI/models/diffusion_models/	base or Turbo (see below)
Krea 2 text encoder (Qwen3-VL)	ComfyUI/models/text_encoders/	load with CLIPLoader type krea2
Krea 2 VAE	ComfyUI/models/vae/

3. Base model — Base vs Turbo

Kroma works on both Krea 2 checkpoints. Turbo is recommended (fewer steps, ~6× faster, and in practice just as good or better with this LoRA):

Checkpoint	Steps	Guidance (CFG)	Shift (mu)
Krea-2-Turbo (recommended)	8	0.0 – 1.0	1.15
Krea-2 base	28	3.5 – 4.5	~1.15

The Turbo sampling values (steps=8, cfg≈0, mu=1.15) are the settings the Turbo checkpoint was distilled for; keep them for best results.

4. Workflow

1. Load Diffusion Model → your Krea 2 (base or Turbo) checkpoint.
2. Load CLIP → the Krea 2 text encoder, set type to krea2.
3. Load VAE → the Krea 2 VAE.
4. LoraLoader (Model only) → select kroma-v0.1.safetensors, strength_model = 1.0. Insert it between the diffusion model loader and the sampler. (The .diff norm / modulation deltas are applied at the same strength; 1.0 reproduces the fine-tune exactly.)
5. KSampler → set steps / CFG / shift per the table above.
6. Prompt and generate.

5. Tips

• Start with strength 1.0; lower it (e.g. 0.7–0.9) if you want a subtler effect.
• If the output looks off, confirm the text encoder is loaded with type krea2 (Krea 2 expects a 12-layer Qwen3-VL stack) and that you're on a current ComfyUI.

---

Conversion / provenance

This LoRA was produced from a K2 (Krea 2) fine-tune checkpoint and converted into the ComfyUI format: LoRA adapters use the native diffusion_model.* key naming, and the fully-fine-tuned RMSNorm / modulation tensors are stored as .diff deltas against the base model. Both are applied by ComfyUI's standard LoraLoader.

License

MIT — see LICENSE. Note that the base Krea 2 model is governed by its own license; this LoRA does not grant any rights to the base weights.