Uncensored AI Image · lodestones
Uncensored Kroma v0.3.1 Turbo OPD: Download, ComfyUI and License
Uncensored Kroma is lodestones' community fine-tune of Krea 2, and Kroma v0.3.1 Turbo OPD is the checkpoint its author currently recommends. This guide covers the download, ComfyUI settings, quantized builds and license terms. If you would rather skip a 26 GB download, you can get close to uncensored Kroma results in the browser with the same base model.
Spicy AI is not Kroma — this page currently runs Seedream, Krea 2 and Wan 2.6.
- Free to start
- No filters
- Runs in browser
Generate an uncensored image
Text to imageDescribe the image you want. Add a reference photo to edit it, or leave it empty for text-to-image.
Try a prompt
Example prompts are in English, which gives the most reliable results.
What is Kroma?
Uncensored Kroma (often searched as Krea Kroma) is a full fine-tune of Krea 2, the open-weight text-to-image model Krea AI trained from scratch and documented in its technical report of June 23, 2026. It is published on Hugging Face by lodestones, the creator of Chroma, and is not a Krea AI product. The base uses a diffusion transformer with a Qwen3-VL text encoder; the base weights are gated and ship as Raw (undistilled) and Turbo (8-step distilled). The fine-tune's repository is tagged Not-For-All-Audiences, its Civitai page is marked NSFW, and the current recommended file is kroma-v0.3.1-turbo-opd.safetensors.
Why people look for an uncensored Kroma
The hosted base model inside Krea's own app is moderated, so people searching for an uncensored Krea model usually land on uncensored Kroma: its teacher checkpoint is named after booru and e6 data, and community workflows are labelled uncensored. But uncensored Kroma does not mean unrestricted. The Community License asks anyone deploying the model to run reasonable content filtering and to follow Krea's Acceptable Use Policy, which bans CSAM, non-consensual intimate imagery, deepfakes of real people and more. Everything you make must depict adults (18+).
Kroma v0.3.1 Turbo OPD download: which file to get
Get uncensored Kroma from the lodestones/Kroma repository on Hugging Face. The file to use is kroma-v0.3.1-turbo-opd.safetensors (26,283,332,880 bytes, about 26.3 GB). The repo also holds kroma-v0.3-base.safetensors (about 51.3 GB, undistilled, meant for training or slow high-quality sampling), a booru/e6-named teacher checkpoint (about 25.7 GB) and a 1.7 GB scanline LoRA. On October 7, 2026 the author deleted the v0.1, v0.2 and v0.3 turbo files, so older tutorials linking the v0.2 turbo file now break, and the README download table still lists that stale name. Civitai hosts the same checkpoint on the author's Civitai model page; Hugging Face is the primary source.
What changed from v0.3 to v0.3.1: on-policy distillation
Earlier turbo builds were distilled offline: the student imitated the teacher on a fixed sampling schedule, which the author says slowly pulls a model off its data manifold and shows up as mode collapse, greyed-out detail and prompts that suddenly stop working. Version 0.3.1 uses on-policy distillation (OPD) instead. The student generates its own trajectories and the teacher corrects it on exactly those states, so training only touches states the model really visits and it stays in its original distribution. The claimed result is turbo speed without the usual distillation tax. That is the publisher's description; no independent benchmark has been published yet.
How to run it in ComfyUI: steps, CFG and shift
To run uncensored Kroma you need a ComfyUI build with native Krea 2 support. Put the checkpoint in ComfyUI/models/diffusion_models/, load the Qwen3-VL text encoder with CLIPLoader set to type krea2, and load qwen_image_vae (both are in the Comfy-Org/Krea-2 repackage). The minimal chain is Load Diffusion Model, KSampler, VAE Decode and Save Image. Recommended sampling is 8 to 12 steps, CFG 1.0 to 1.5, euler with the simple scheduler and shift 1.15. If outputs look broken, first check the encoder type and your ComfyUI version.
VRAM requirements: INT8 ConvRot vs w6a8 vs w4a8 vs GGUF
The bf16 uncensored Kroma checkpoint alone is about 26.3 GB, more than most consumer cards hold. A third-party repo by jakethessmake, which states it is not a Krea or lodestones release, offers ComfyUI-native quants: int8 ConvRot at about 13 GB (self-reported 1.0% mean relative weight error), w6a8 at about 10 GB (2.1%) and w4a8 at about 7 GB (7.3%), ranked in that order for quality; w6a8 and w4a8 need a comfy-kitchen version with those formats. Abiray publishes GGUF files from Q2_K (about 6 GB target) to Q6_K (16 GB and up), with Q4_K_M suggested as the daily driver for 12 GB cards. Real usage also depends on resolution and the text encoder, so test on your own hardware.
Kroma vs Krea 2 Turbo, Wulver and v0.3
Krea 2 Turbo is Krea AI's own 8-step distilled release: the best choice for its intended look and the cleanest license story. Uncensored Kroma keeps the same text encoder and VAE but adds its own training and a distinct style, so prompts do not transfer one to one. Wulver is another full fine-tune of the Raw base that targets anthro and furry art, a narrower niche. Against v0.3 turbo, v0.3.1 swaps offline distillation for OPD at the same speed class, and the old v0.3 turbo file is no longer in the repo.
LoRA training and loading
The author says that because OPD keeps the student inside its original distribution, LoRAs trained the normal way load directly onto uncensored Kroma (v0.3.1-turbo-opd). Train against the teacher checkpoint or v0.3-base rather than the turbo file. The community also shares a rank-384 delta LoRA by Hiccup1234 (2.8 GB) that moves the base model toward uncensored Kroma v0.3.1; its author reports it keeps about 38% of the delta energy, so treat it as an approximation you blend by strength.
License and commercial use
The uncensored Kroma license is not fully settled. The Hugging Face metadata names the Krea 2 Community License, while the README text says MIT and points to a LICENSE file that is not in the repo. Because the weights derive from Krea's model, plan around its Community License: commercial use is allowed only while your company-wide annual revenue stays under 1,000,000 USD, above that you need an enterprise license from Krea, and redistributed derivatives must include the agreement, keep the notice, start the model name with Krea and must not claim to be an official Krea product. Check the final license text before any commercial use.
What you can create
Adult boudoir and glamour
Soft window light, film grain and intimate poses featuring consenting adult subjects.
Illustration and anime
The booru-trained teacher gives uncensored Kroma a strong pull toward anime and illustrated styles.
Fantasy pin-ups
Succubi, sorceresses and dark-fantasy characters, always depicted as adults.
Prompt tips for Kroma-style results
- Write natural sentences; the Qwen3-VL encoder reads full descriptions better than tag lists.
- Stay at CFG 1.0 to 1.5 and 8 to 12 steps; higher CFG tends to burn turbo outputs.
- Expect uncensored Kroma to look different from its base and adjust prompts instead of copying them one to one.
- Add a light subject LoRA when you need a consistent character style.
- State the subject's adult age in prompts and never reference real people.
Self-hosting Kroma vs generating online on Spicy AI
| Local deployment (official weights) | Spicy AI online | |
|---|---|---|
| Content filters | Set by lodestones | No content filters |
| Open weights | Yes | Not needed — runs in the browser |
| Model | Kroma | Seedream, Krea 2 and Wan 2.6 |
| Price to start | Varies by plan | Free tier available |
- Text-to-image from a prompt
- Image-to-image edits from a reference photo
- Up to 4 images per run in multiple aspect ratios
FAQ
Does Spicy AI run uncensored Kroma v0.3.1 Turbo OPD?
No. The fine-tune is not hosted on Spicy AI. The image generator here uses the base model uncensored Kroma was trained from, plus Seedream and Wan 2.6, so results look different from a local run of the checkpoint.
Is the jakethessmake quant repo an official release?
No. Its own notice says it is not a Krea or lodestones release; it only changes the number format of the weights and does no retraining.
How many parameters does it have?
Hugging Face metadata for the base model totals about 12.82B parameters, but that count includes the text encoder and other components. No transformer-only figure has been published, so file size is not a parameter count.
Is Spicy AI affiliated with lodestones?
No. Spicy AI is an independent platform and is not affiliated with or endorsed by lodestones. Kroma is a trademark of its respective owner.
Is it free?
You can start free with daily generations, and buy credits when you need more.
Sources
Spicy AI is not affiliated with lodestones. Kroma and other product names are trademarks of their respective owners and are used here for identification only. All content must depict adults (18+).