Chroma is an 8.9B parameter, Apache 2.0 licensed de-distilled fork of Flux.1 Schnell that is uncensored by design. It runs in ComfyUI, needs roughly 12GB or more VRAM (less with GGUF quants), takes natural-language prompts like Flux, and produces adult content without the safety alignment that limits the base Flux models.
The Flux family gave the open community stunning image quality, but the official releases are aligned to avoid explicit content, and Schnell in particular is distilled in a way that resists fine-tuning. Chroma is the community’s answer. It is a de-distilled, retrained Flux fork built from the ground up to be unlocked and trainable, released under a permissive Apache 2.0 license. This guide explains what it is, why it is uncensored, how to install and configure it in ComfyUI, VRAM and GGUF options, the right sampler and step settings, how to prompt it, and where it falls short. Every subject discussed is a fictional adult over 18.
What Chroma is and why it is uncensored
Chroma starts from Flux.1 Schnell, the fast distilled member of Black Forest Labs’ Flux family. Two things make it distinct.
First, it is de-distilled. Schnell was distilled for few-step speed, which also bakes in behavior that makes it hard to fine-tune. Chroma reverses that, restoring a model that responds properly to training and guidance. That is what allowed the community to retrain it on a broad dataset instead of the filtered one behind the official release.
Second, it is uncensored by design. Where the official Flux checkpoints are aligned to refuse or degrade explicit content, Chroma was retrained specifically to remove that limitation. It understands adult anatomy and NSFW concepts natively, without needing a stack of corrective LoRAs just to get the base model to cooperate. That is the core reason people choose it over fighting base Flux.
The Apache 2.0 license matters too. It is genuinely permissive, so you can use, modify, and build on Chroma freely. That openness is why an ecosystem of quants and fine-tunes formed around it quickly.
One more thing worth knowing about the de-distillation is what it does to output character. Because Chroma was retrained rather than merely unlocked, it does not behave identically to base Flux; it has its own aesthetic and its own handling of anatomy and detail. People coming from official Flux sometimes expect a pixel-for-pixel match with the safety removed, but that is not what Chroma is. It is a genuinely retrained model in the Flux family, so treat its look as its own and prompt to it directly rather than assuming your old Flux prompts transfer unchanged.
This makes Chroma distinct from our general Flux NSFW guide, which covers the base Flux models and the workarounds needed to coax adult output from them. Chroma is the purpose-built uncensored member of that same family, so if you have read that guide, think of this as the version where you skip the fight.
Where to download Chroma
Chroma is hosted on Hugging Face. You want the main model weights as a .safetensors file, and you will also need the Flux text encoders (the CLIP and T5 encoders that the Flux architecture uses) and the Flux VAE. Many workflow bundles package the encoders and VAE together with instructions.
Folder placement in ComfyUI:
- Main Chroma model goes in
ComfyUI/models/unet/ordiffusion_models/. - The CLIP and T5 text encoders go in
ComfyUI/models/clip/ortext_encoders/. - The Flux VAE goes in
ComfyUI/models/vae/.
If you also want GGUF quantized versions for lower VRAM, those live in the same or a dedicated GGUF repository. More on quants below.
ComfyUI setup
Chroma runs in ComfyUI, the node-based interface that gives you full control over the generation graph. If you are new to it, start with our ComfyUI for NSFW AI complete guide, which covers installation, the manager, and loading workflows.
For Chroma specifically:
- Update ComfyUI to a recent version so it has current Flux and Chroma node support.
- Place the model, text encoders, and VAE in the folders listed above.
- Load a Chroma or Flux workflow template. The community publishes ready-made graphs, and using one saves you from wiring the dual text encoder setup by hand.
- Point the loader nodes at the Chroma model, the CLIP and T5 encoders, and the Flux VAE.
- Confirm the workflow uses the Flux-style sampling nodes rather than the classic SD ones.
Because Chroma inherits the Flux architecture, it uses the Flux sampling path, not the older Stable Diffusion one. Using a mismatched workflow is the most common reason a first run produces garbage.
VRAM and GGUF quantization
At 8.9B parameters, the full-precision Chroma model is not tiny. Expect to want roughly 12GB or more VRAM for a comfortable full-weight experience. That puts the unquantized model out of reach for smaller cards, which is exactly what GGUF quantization solves.
GGUF is a quantization format that shrinks the model by storing weights at lower precision. You trade a small amount of quality for a large drop in VRAM use. Community GGUF builds of Chroma come in several sizes so you can pick the largest that fits your card.
| VRAM | Recommended build | Notes |
|---|---|---|
| 16GB and up | Full or high-bit GGUF | Best quality, full speed |
| 12GB | Full weights or Q8 GGUF | Comfortable, near-full quality |
| 8GB | Q5 or Q6 GGUF | Solid results, small quality trade |
| 6GB | Q4 GGUF with offloading | Works but slower, some quality loss |
To run GGUF you need the GGUF loader custom node in ComfyUI, then you load the quantized file in place of the full model. Everything else in the workflow stays the same. For a broader look at running heavy models on limited hardware, our best NSFW checkpoints for low VRAM guide is a useful companion, and AMD owners should see the Stable Diffusion AMD GPU guide.
Sampler, steps, and CFG
Chroma is de-distilled, which is important for settings. Base Flux Schnell is a few-step model, but because Chroma was de-distilled it behaves more like a normal, trainable diffusion model. That means it generally wants more steps than raw Schnell and responds to real guidance.
A sensible starting configuration:
| Setting | Starting value |
|---|---|
| Steps | 20 to 30 |
| Guidance / CFG | modest, follow the workflow default |
| Sampler | euler or the Flux-recommended sampler |
| Scheduler | simple or the workflow default |
| Resolution | 1024×1024 or Flux-supported sizes |
Flux-family models use a guidance mechanism that is not identical to classic SD CFG, so follow the values baked into a known-good Chroma workflow and adjust from there rather than importing SDXL numbers. If images look undercooked, add steps. If they look over-processed, ease the guidance. Our NSFW AI CFG and sampler settings guide explains how these knobs interact if you want the full theory.
How to prompt Chroma
Chroma prompts like Flux: natural language works well. Instead of a pile of comma-separated tags, you can write descriptive sentences. Describe your fictional adult subject, the setting, the lighting, the mood, the camera, and the style in plain language, as if briefing a photographer.
Because it is uncensored by design, you do not need elaborate jailbreak phrasing to get adult output. State what you want clearly and tastefully. The model understands the concepts natively, so clarity beats trickery.
A practical structure:
- Lead with the subject and the core action or pose.
- Add physical and wardrobe detail.
- Set the environment and lighting.
- Finish with style and camera language (photographic, film look, soft light, and so on).
Flux-family models tend to reward coherent, well-written prompts and can struggle with contradictory instructions, so keep the description internally consistent.
The baseline safety negative prompt
Even on an uncensored model, always keep a safety negative prompt that excludes disallowed content. Include terms such as child, minor, underage, loli, and shota. Every subject must be a clearly adult, fictional person over 18. This is a hard rule across every model in this series. Note that Flux-style workflows sometimes handle negatives differently from classic SD, so make sure your workflow actually applies the negative prompt rather than ignoring it, and keep those safety terms in place regardless.
Strengths of Chroma
- Uncensored natively. No LoRA stack just to unlock the base model. Adult concepts are understood out of the box.
- Flux-level quality. It inherits the Flux family’s strong coherence, anatomy, and lighting.
- Trainable. Being de-distilled, it fine-tunes and accepts LoRAs properly, unlike raw Schnell.
- Permissive license. Apache 2.0 means you can build on it freely.
- Natural language prompting. Easy for people who dislike tag soup.
If you want to specialize it further, you can train a LoRA on top. Our how to train a NSFW LoRA guide walks through the process, and Chroma’s trainability makes it a good target compared to models that resist fine-tuning.
A prompt walkthrough for Chroma
Because Chroma prompts in natural language, the best way to learn it is to build a prompt like a photographer’s brief. Start with a plain sentence naming your fictional adult subject and what they are doing. Add a sentence of physical and wardrobe detail. Add a sentence setting the scene and the time of day. Add a sentence on lighting and mood. Finish with the visual style, whether that is a soft photographic look, a film grain aesthetic, or a particular color palette.
Written this way, a Chroma prompt reads like prose, not a tag list, and the Flux-derived architecture handles it beautifully. Two habits pay off. First, keep the description internally consistent, since Flux-family models get confused by contradictory instructions (for example asking for both harsh noon sun and soft candlelight in the same image). Second, because Chroma is uncensored by design, you do not need coded language or jailbreak phrasing. Say plainly and tastefully what you want, and the model understands it natively.
If a generation is close but off, adjust one clause and regenerate rather than rewriting everything. Change the lighting sentence, or the camera sentence, or ease the guidance slightly. Single-variable iteration teaches you how Chroma responds far faster than wholesale rewrites.
Finishing and upscaling
Chroma generates at Flux-supported resolutions and benefits from a finishing pass. Generate candidates, pick the keepers, then upscale only those to save compute. Our NSFW AI hires fix complete guide covers the denoise-and-upscale loop, and the NSFW AI upscaler guide compares dedicated upscale models for the final export. Keep the guidance moderate on the upscale so the second pass adds detail without over-processing the image you already approved.
If you are running a GGUF quant on a tighter card, the upscale pass is the most VRAM-hungry moment of the workflow, so close other GPU applications first and upscale in smaller steps if you hit memory limits.
Limitations to expect
- VRAM. At 8.9B the full model wants 12GB or more. GGUF helps but adds a small quality cost.
- Speed. Because it is de-distilled and runs more steps than Schnell, it is slower than few-step turbo models like Z-Image Turbo. If speed is your priority, see the Z-Image Turbo guide.
- Flux workflow specifics. The dual text encoder and Flux sampling path trip up newcomers. Use a template.
- Not a tag model. If you are used to Danbooru-style tag prompting on anime SDXL forks, the natural-language style takes adjustment.
For problems like black images, missing text encoder errors, or a workflow that outputs noise, the NSFW AI troubleshooting guide and the ComfyUI not working fix cover the common failure modes, most of which come down to a wrong VAE, a missing encoder, or an SD workflow used on a Flux model.
Chroma versus other options
Chroma is the best pick when you want Flux-grade image quality with native uncensored behavior and the ability to fine-tune. If you would rather not run 8.9B locally at all, the various SDXL forks remain lighter and faster for many adult use cases; see the best Stable Diffusion checkpoints for NSFW roundup. And if you want to test drive high quality NSFW generation with zero setup before committing to a local Chroma install, the free NSFW AI generator on this site runs in your browser with no login, which is a quick way to sanity check your prompt ideas before wiring up a full ComfyUI graph.
Common first-run problems
Most Chroma first-run failures come down to three things. Using an old Stable Diffusion workflow on a Flux-architecture model produces pure noise, so always load a Chroma or Flux template that uses the Flux sampling path. Forgetting one of the two text encoders (the CLIP and the T5) causes load errors or garbage output, so confirm both are in place. And loading the wrong VAE gives washed-out or wrongly colored images, so use the Flux VAE specifically.
If you are on a GGUF build and it will not load, the usual cause is a missing GGUF loader custom node, which you install through the ComfyUI manager. Once the architecture, encoders, VAE, and (if needed) GGUF loader are all correct, Chroma runs reliably. The ComfyUI not working fix covers these node and loader errors in detail.
First-run checklist
- Download Chroma weights, the Flux text encoders, and the Flux VAE from Hugging Face.
- If your VRAM is under 12GB, grab a GGUF quant that fits and install the GGUF loader node.
- Update ComfyUI and load a known-good Chroma or Flux workflow.
- Point the loaders at the model, encoders, and VAE.
- Set steps to 20 to 30 and use the workflow’s guidance and sampler defaults.
- Write a clear, natural-language prompt describing your fictional adult subject.
- Keep the safety negative prompt in place at all times.
- Generate, then tune steps and guidance one at a time.
Chroma is the standout choice in 2026 for anyone who loved Flux image quality but hit its content wall. It is uncensored by design, properly trainable, permissively licensed, and it prompts in plain English. Get the Flux workflow right, pick a build that fits your VRAM, and it delivers some of the best open uncensored output available.
Frequently asked questions
Is Chroma really uncensored out of the box?
Yes. Chroma was de-distilled from Flux.1 Schnell and retrained specifically to remove the safety alignment that limits the official Flux models. It understands adult anatomy and NSFW concepts natively, so you do not need a stack of corrective LoRAs just to get the base model to produce explicit content the way you do with base Flux.
How much VRAM does Chroma need?
The full 8.9B model wants roughly 12GB or more for a comfortable experience. GGUF quantized builds lower that substantially: a Q5 or Q6 GGUF runs on 8GB, and a Q4 build with offloading can work on 6GB. You trade a small amount of quality for a large drop in VRAM usage with quants.
Do I need ComfyUI for Chroma?
ComfyUI is the standard interface for Chroma because it inherits the Flux architecture and needs the Flux sampling path plus dual text encoders. Use a ready-made Chroma or Flux workflow template rather than wiring it yourself. If you only want to test prompts quickly, this site’s free browser generator lets you try NSFW generation with no install first.
What is the difference between Chroma and base Flux?
Base Flux models are aligned to avoid explicit content, and Schnell is distilled in a way that resists fine-tuning. Chroma is a de-distilled fork of Schnell that was retrained to be uncensored and trainable, released under Apache 2.0. In short, Chroma is the Flux-family model built for adult content and LoRA training.
What sampler and steps work best for Chroma?
Because Chroma is de-distilled it behaves like a normal diffusion model, so use around 20 to 30 steps with euler or the Flux-recommended sampler and the guidance value baked into a known-good workflow. Do not import SDXL CFG numbers, since Flux-family guidance works differently. Add steps if output looks undercooked.
Is Chroma free?
Yes. Chroma is released under the permissive Apache 2.0 license and the weights are free on Hugging Face, so you can download, use, modify, and build on it at no cost. The only expense is your own hardware and power for local generation, or you can use this site’s free browser generator instead.
Can I train a LoRA on Chroma?
Yes, and this is one of its strengths. Because Chroma was de-distilled, it fine-tunes and accepts LoRAs properly, unlike raw Flux Schnell which resists training. That makes it a strong base for building style or content LoRAs. Our NSFW LoRA training guide covers the full workflow if you want to specialize its output further.
How does Chroma prompt compared to SDXL anime models?
Chroma prompts in natural language like the rest of the Flux family, so you write descriptive sentences rather than comma-separated Danbooru tags. If you come from tag-based anime SDXL forks, this takes adjustment, but it rewards clear, coherent, well-written descriptions and tends to struggle only when a prompt contains contradictory instructions.
Compare the options:
