Kroma v0.3: Krea 2 Fine-Tune Adds Full Base Checkpoint
Lodestones ships Kroma v0.3 for Krea 2: a non-distilled full base checkpoint, a turbo variant, and a community LoRA pack that unlocks prompt following fixes.
Lodestones uploaded Kroma v0.3 to the Kroma repo on Hugging Face between August 19 and 30, the third iteration of the community fine-tune of Krea 2. v0.3 comes in two weights: a full base checkpoint (about 51.3 GB) that has no distilled turbo delta baked in, and the familiar turbo checkpoint (about 25.6 GB) for fast sampling. A separate kroma-v0.3-tdm-4steps-artifact file (about 234 MB) rounds out the release.
What is new in v0.3
The headline change is that Kroma now exists as a raw-base fine-tune, not just a turbo-merged checkpoint. Krea 2 ships in two flavors: the raw (non-distilled) model that produces the highest quality but needs more steps, and the Turbo distillation that samples in 8-12 steps. Kroma v0.1 and v0.2 could only be used at Turbo speed, because the fine-tune was trained on (or merged with) the Turbo delta. With v0.3-base, the fine-tune finally covers the non-distilled weights, so the "no distillation artifacts" route is open for Kroma too.
| File | Size | What it is |
|---|---|---|
kroma-v0.3-base.safetensors | ~51.3 GB | Full fine-tune of Krea 2 raw (non-distilled) weights |
kroma-v0.3-turbo.safetensors | ~25.6 GB | Turbo-speed variant, sampled like Krea 2 Turbo |
kroma-v0.3-tdm-4steps-artifact.safetensors | ~234 MB | Experimental 4-step TDM artifact |
Community extras already circulating
The Banodoco krea-2-image channel moved fast on the release:
- INT8 ConvRot quants of the v0.3 turbo checkpoint appeared within hours from Hiccup1234 and thedarkthrust, with a cutecaption collection tracking the conversions.
- Silveroxides maintains Kroma-LoRA with rank-384 LoRA extracts, including a
custom-unlockbuild for v0.3 base. - Hippotes published a
kroma-v0.3-txtfusion_diffextraction on Krea-2-Experiments that lets you take just the text-fusion layers of Kroma and stack them on the official Krea 2 Turbo checkpoint, which some users report as the best of both worlds.
How to use Kroma v0.3 in ComfyUI
The fine-tune uses the standard Krea 2 stack, so it works with ComfyUI's native Krea 2 support:
- Download the checkpoint into
ComfyUI/models/diffusion_models/ - Load the Krea 2 text encoder (Qwen3-VL, 12 layers) with a CLIPLoader set to type
krea2 - Load the Krea 2 VAE
- For
kroma-v0.3-turbo: sample with Turbo settings (8-12 steps, CFG 1.0-1.5, shift/mu 1.15). Forkroma-v0.3-base: use regular sampling settings as you would for the Krea 2 raw checkpoint.
One caveat from early users: Kroma is a fine-tune with its own "vibe", and prompts do not always translate 1:1 from Krea 2. If a subject stops following the prompt, a small amount of a subject LoRA nudges it back.
Availability
- Kroma v0.3 on Hugging Face (MIT license on the fine-tune; Krea 2 itself keeps its own community license)
- Community INT8 quant collection
Comments
Sign in with GitHub to join the discussion.