# Chroma
Chroma (Chroma1-Base) is an 8.9-billion-parameter, open-weights text-to-image model derived from FLUX.1-schnell and released in August 2025 under the Apache 2.0 license by a developer working under the pseudonym Lodestone Rock, with training funded by the community rather than a company.[1](https://huggingface.co/lodestones/Chroma1-Base/blob/main/README.md) • [2](https://huggingface.co/lodestones/Chroma/raw/main/README.md) • [3](https://nomusica.com/chroma1-launches-as-open-source-alternative-to-midjourney-and-nano-banana/) It is framed by its creator and ecosystem sites as a fully open community alternative to FLUX, whose main quality-tier model carries a non-commercial license.
| Fact | Value |
|---|---|
| Parameters | 8.9B (down from FLUX.1-schnell's 12B)[1](https://huggingface.co/lodestones/Chroma1-Base/blob/main/README.md) • [2](https://huggingface.co/lodestones/Chroma/raw/main/README.md) |
| Base architecture | FLUX.1-schnell, de-distilled[2](https://huggingface.co/lodestones/Chroma/raw/main/README.md) |
| License | Apache 2.0, full commercial use[1](https://huggingface.co/lodestones/Chroma1-Base/blob/main/README.md) |
| Training | ~105,000 H100 GPU hours on 5M images curated from a 20M pool[4](https://civitai.com/models/1330309/chroma?modelVersionId=1884429) • [1](https://huggingface.co/lodestones/Chroma1-Base/blob/main/README.md) |
| Text encoder | T5 (t5-v1_1-xxl per ecosystem descriptions)[5](https://civitai.com/ecosystems/chroma) |
| Release | August 2025, pseudonym Lodestone Rock, community-funded[3](https://nomusica.com/chroma1-launches-as-open-source-alternative-to-midjourney-and-nano-banana/) |
| Family | Chroma1-Base, Chroma1-HD, Chroma1-Flash, Chroma1-Radiance (WIP)[4](https://civitai.com/models/1330309/chroma?modelVersionId=1884429) |
| Adoption | 56.9K Chroma1-HD downloads on Hugging Face as of early August 2026; 516+ LoRAs on Civitai[6](https://localaimaster.com/blog/chroma-local-guide) • [5](https://civitai.com/ecosystems/chroma) |
## Architecture and training as published
Chroma is neither a fine-tune nor a from-scratch model; it is a surgical modification of FLUX.1-schnell followed by a retrain. According to the developer's architecture notes, FLUX carries 3.3B parameters whose only job is to encode the current denoising timestep together with pooled CLIP vectors, in effect 8 bytes of float values. A zeroing experiment showed the pooled CLIP vectors contributed little to the output, so the developer replaced that block with a 250M-parameter feed-forward network; the replacement itself took one day on a single RTX 3090 and reduced the model from 12B to 8.9B parameters.[2](https://huggingface.co/lodestones/Chroma/raw/main/README.md)
The retrain used a 5M-sample dataset curated from a 20M pool, covering artistic, photographic, and niche styles including anime and furry content. The developer describes the model as fully uncensored, "reintroducing missing anatomical concepts" removed from FLUX's training.[2](https://huggingface.co/lodestones/Chroma/raw/main/README.md) The base training run consumed about 105,000 H100 GPU hours, funded by the community.[4](https://civitai.com/models/1330309/chroma?modelVersionId=1884429) The creator states that no aesthetic tuning or post-training such as DPO was applied, positioning the checkpoints as raw, neutral starting points for fine-tuning.[4](https://civitai.com/models/1330309/chroma?modelVersionId=1884429) A comprehensive technical report on the modifications and training process was still forthcoming as of the model card.[1](https://huggingface.co/lodestones/Chroma1-Base/blob/main/README.md)
## Licensing, availability and local hardware
Apache 2.0 is the release's central distinction: it grants anyone the right to use, modify, and build on the model commercially.[1](https://huggingface.co/lodestones/Chroma1-Base/blob/main/README.md) FLUX.1-dev, the higher-quality FLUX variant, is restricted to non-commercial use, and FLUX.1-schnell is a distilled model locked to a 4-step generation process. Chroma's release announcement states that "everything is fully Apache 2.0 licensed, no gatekeeping."[4](https://civitai.com/models/1330309/chroma?modelVersionId=1884429)
The weights are hosted on [Hugging Face](https://www.edgepedia.org/hugging-face) as Chroma1-Base and Chroma1-HD.[1](https://huggingface.co/lodestones/Chroma1-Base/blob/main/README.md) Chroma1-HD ships as a 17.8GB BF16 checkpoint that needs roughly 24GB of VRAM to run in full precision; community GGUF quantizations fit smaller cards, with 9.74GB (Q8_0) for 16GB cards, 7.65GB (Q6_K) for 12GB, and 5.57GB (Q4_K_M) for 8GB. It integrates with ComfyUI workflows.[6](https://localaimaster.com/blog/chroma-local-guide)
## Capabilities and benchmarks: claims versus measurement
The creator and community reports describe strong realistic human anatomy, handling of complex prompts, and true classifier-free guidance with [negative prompts](https://www.edgepedia.org/negative-prompt) at 26 to 40 steps.[3](https://nomusica.com/chroma1-launches-as-open-source-alternative-to-midjourney-and-nano-banana/) • [6](https://localaimaster.com/blog/chroma-local-guide)
The speed claims conflict. NoMusica reported community benchmarks claiming Chroma1 can operate up to 2.5 times faster than comparable models on hardware such as the RTX 3080.[3](https://nomusica.com/chroma1-launches-as-open-source-alternative-to-midjourney-and-nano-banana/) The LocalAI Master guide reaches the opposite conclusion from the architecture itself: de-distillation is "the feature and the bill," since Chroma needs 26 to 40 real steps against FLUX.1-schnell's distilled 4, and the guide states plainly that it is not fast and that it published no benchmarks of its own.[6](https://localaimaster.com/blog/chroma-local-guide)
## Comparison with FLUX, SDXL and Qwen-Image
Civitai's editorial comparison positions the models by purpose rather than by a single ranking: FLUX.1 leads on prompt polish and in-image text, [SDXL](https://www.edgepedia.org/sdxl) offers the deepest LoRA library and versatile realism, [Qwen-Image](https://www.edgepedia.org/qwen-image) is stronger at legible typography, and Chroma's case is openness plus a clean, un-tuned fine-tuning base.[5](https://civitai.com/ecosystems/chroma) The LoRA ecosystem numbers make the gap concrete: 516+ Chroma LoRAs against 38K+ each for FLUX.1 and SDXL and 1K+ for Qwen-Image.[5](https://civitai.com/ecosystems/chroma) No direct head-to-head measurements against [Stable Diffusion 3](https://www.edgepedia.org/stable-diffusion-3).5 or Qwen-Image on prompt adherence, typography, or anatomy were found; the comparison is qualitative.
## Reception, ecosystem and adoption
Development was community-funded, and community forums on Reddit and Hugging Face actively shaped post-release work.[3](https://nomusica.com/chroma1-launches-as-open-source-alternative-to-midjourney-and-nano-banana/) Adoption is measurable but modest against the incumbents: Chroma1-HD had 56.9K downloads on Hugging Face as of early August 2026, and Civitai listed 516+ LoRAs.[6](https://localaimaster.com/blog/chroma-local-guide) • [5](https://civitai.com/ecosystems/chroma)
The family grew through 2026. Chroma1-HD is the finished, stable 1024x1024 high-resolution fine-tune of the 512x512-core Base. Chroma1-Flash is an experimental speed fine-tune exploring how to accelerate flow-matching models without GAN training, with transferable delta weights. Chroma1-Radiance is a work-in-progress pixel-space variant intended to avoid VAE compression artifacts. Later experiments include Chroma2-[Kaleidoscope](https://www.edgepedia.org/kaleidoscope), a WIP fine-tune of FLUX.2-klein-base-4B, and an early project called Kroma. All are Apache 2.0.[4](https://civitai.com/models/1330309/chroma?modelVersionId=1884429) • [6](https://localaimaster.com/blog/chroma-local-guide)
## Controversies and criticisms
The model card states that Chroma was not aligned with a safety filter, may generate harmful, explicit or offensive content, and that responsibility for safeguards rests with downstream developers; the training data was deliberately uncensored.[1](https://huggingface.co/lodestones/Chroma1-Base/blob/main/README.md) • [2](https://huggingface.co/lodestones/Chroma/raw/main/README.md) This is a disclosed design choice rather than a discovered lapse, but it places the model in the uncensored-image-generation category by intent.
The "fully open" framing is accurate at the license level, yet the performance story is thinner than the marketing: the 2.5x-faster claim contradicts the 26-to-40-step cost of de-distillation, and no independent benchmark supports the quality claims.[3](https://nomusica.com/chroma1-launches-as-open-source-alternative-to-midjourney-and-nano-banana/) • [6](https://localaimaster.com/blog/chroma-local-guide) No training-data provenance disputes, governance conflicts, or benchmark-gaming allegations appear in the sources.
## References
1. [README.md · lodestones/Chroma1-Base (official model card)](https://huggingface.co/lodestones/Chroma1-Base/blob/main/README.md)
2. [README.md · lodestones/Chroma (developer architecture notes)](https://huggingface.co/lodestones/Chroma/raw/main/README.md)
3. [Free AI Image Generator Chroma1 Released by Independent Creator | NoMusica.com](https://nomusica.com/chroma1-launches-as-open-source-alternative-to-midjourney-and-nano-banana/)
4. [Chroma - v1.0-HD | Chroma Checkpoint (Civitai release announcement)](https://civitai.com/models/1330309/chroma?modelVersionId=1884429)
5. [Chroma AI Models & Generator | Civitai ecosystem page](https://civitai.com/ecosystems/chroma)
6. [Chroma Local Guide: The Apache-2.0 Uncensored FLUX Model](https://localaimaster.com/blog/chroma-local-guide)
---
*Topic: Encyclopedia › Technology and the built world › Computing and digital systems › Modern AI: foundation models, generative AI, and the AI industry › Model families and named models › Image generation models*
*Initially written Sep 17, 2026 · Reviewed: Sep 30, 2026 · Edited: Oct 11, 2026 · Last review: Sep 30, 2026*
*Copyright 2026 EdgeChat AI, a subsidiary of Biostate AI.*
License: Edgepedia Community License 1.0, https://www.edgepedia.org/edgepedia/license