Kroma v0.3.1 OPD — recommended
Use kroma-v0.3.1-turbo-opd.safetensors. This is the current recommended
checkpoint.
What OPD is: v0.3.1 was distilled on-policy. Instead of the classic offline recipe — imitating the teacher on a fixed sampling schedule, which slowly pulls the student off the original data manifold — the student generates its own trajectories and the teacher corrects it on those exact points. Training only ever happens on states the model actually visits, so the distilled model stays inside the original model's distribution: Turbo speed without the usual distillation tax (mode collapse, washed-out detail, prompts that suddenly stop working).
Running it in ComfyUI is identical to v0.2 — same Krea 2 text encoder (CLIPLoader
type krea2), same VAE, same Turbo sampling settings.
LoRA training
Because OPD keeps the model within the original distribution, LoRAs trained the
normal way carry straight over. Train against either the teacher
(kroma-sensei-booru-e6-teacher-velocity-v0.3.safetensors) or the base
(kroma-v0.3-base.safetensors) — the resulting LoRA loads on
kroma-v0.3.1-turbo-opd.safetensors unchanged.
Running in ComfyUI
1. Requirements
- A recent ComfyUI with native Krea 2 support (update to the latest master).
- The Krea 2 text encoder and VAE (the diffusion model is this repo — no separate base checkpoint needed).
2. Files to download
Place these in your ComfyUI folders:
| What | Where | Notes |
|---|---|---|
kroma-v0.2-turbo.safetensors |
ComfyUI/models/diffusion_models/ |
this repo |
| Krea 2 text encoder (Qwen3-VL) | ComfyUI/models/text_encoders/ |
load with CLIPLoader type krea2 |
| Krea 2 VAE | ComfyUI/models/vae/ |
3. Sampling settings
The Krea 2 Turbo delta is merged in, so sample with Turbo settings:
| Setting | Value |
|---|---|
| Steps | 8-12 |
| Guidance (CFG) | 1.0 – 1.5 |
| Shift (mu) | 1.15 |
These are the settings the Turbo checkpoint was distilled for; keep them for best results.
4. Workflow
- Load Diffusion Model → select
kroma-v0.2-turbo.safetensors. - Load CLIP → the Krea 2 text encoder, set type to
krea2. - Load VAE → the Krea 2 VAE.
- KSampler → steps / CFG / shift per the table above.
- Prompt and generate.
Minimal node chain:
Load Diffusion Model (kroma-v0.2) ─► KSampler ─► VAE Decode ─► Save Image
Load CLIP (krea2) ─────────────────► (clip) ▲
Load VAE ───────────────────────────────────────┘
Empty Latent / prompt ────────────────────────────►
5. Tips
- 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.
Provenance
v0.2 was produced by continued full fine-tuning of the K2 (Krea 2) checkpoint behind Kroma v0.1 — without the rank-256 LoRA delta compression used for the v0.1 release — followed by merging the Krea 2 Turbo delta in as a rank-512 LoRA.
License
MIT — see LICENSE. Note that Krea 2 is governed by its own license; this
fine-tune does not grant any rights to the base weights.