Chroma
Chroma is an open-weight, Apache-2.0 text-to-image model by Lodestone — a true base model with no aesthetic tuning, meant as a raw, neutral foundation for fine-tuning. Built as a de-distilled Flux.1 Schnell (~8.9B parameters) with a T5 text encoder, it reads natural-language prompts and stays deliberately un-opinionated. Generate with the Civitai-hosted Chroma checkpoint right here — no GPU, no install.
About Chroma
Chroma is an open-source foundational image model from Lodestone, released fully under Apache 2.0 with no gatekeeping. It is positioned explicitly as a true base model: no aesthetic tuning and no post-training such as DPO, so it ships as a raw, neutral starting point designed for the community to fine-tune rather than a polished, ready-styled generator. Architecturally it is a de-distill of Flux.1 Schnell at roughly 8.9 billion parameters, pairing the transformer with a T5 text encoder that actually parses grammar — so it reads natural-language descriptions rather than comma-separated tags. The base-model training run consumed about 105,000 H100 hours, packing in a broad data distribution intended to make fine-tuning on top of it converge quickly.
Chroma is really a small family rather than one file. Chroma1-Base is the core 512×512 foundation, aimed at longer fine-tunes; Chroma1-HD is the 1024×1024 high-res fine-tune of that base and the best starting point for high-res LoRAs — it is the version hosted on Civitai. On the research branch, Chroma1-Flash is an experimental speed fine-tune whose delta weights can be applied to other Chroma versions, and Chroma1-Radiance is a work-in-progress pixel-space variant meant to sidestep VAE compression artifacts. For local users there are FP8-scaled and GGUF quantized builds that lower the VRAM needed to run it.
Choose Chroma when you want a permissive, un-tuned open foundation to build on — a neutral canvas for fine-tunes and LoRAs rather than a model with a baked-in house style. Because it is a Flux-derived model with a T5 encoder, it follows natural-language prompts well, but it will not hand you the polished, aesthetic-tuned look that a post-trained model like Flux.1 Dev produces out of the box; that is by design. For maximum prompt polish and in-image text, Flux leads; for the deepest LoRA library and versatile realism, SDXL is the safer default; for legible typography, Qwen-Image is stronger. Reach for Chroma when openness, a clean license, and a raw base to fine-tune are what you actually need.
How to prompt Chroma
- Write in natural language — full descriptive sentences, not comma-separated tags. Chroma’s T5 text encoder parses grammar, so tag-style prompts underperform. A useful template is: [subject with detail], [setting/scene], [style and color palette], [lighting], [composition].
- Give the encoder enough to work with. Very short prompts tend to underperform on Chroma — add concrete detail about materials, mood, and framing rather than leaving it terse.
- Skip weight syntax like (word:1.4) — it is not used here. Control emphasis through descriptive language and word choice instead of numeric weights.
- Use a negative prompt — Chroma supports it as a separate parameter and benefits from one. A solid quality-focused baseline is "low quality, ugly, unfinished, out of focus, deformed, disfigured, blurry, flat colors."
- Chroma uses true CFG (classifier-free guidance). The official example uses a guidance scale around 3.0; some nudge it a little higher for punchier output — worth trying both when dialing in a prompt.
Featured Chroma models
Curated — the models worth generating with first.
Popular Chroma LoRAs & add-ons
Top LoRAs by downloads — live data, refreshed daily. Stack them on any checkpoint.
Example generations
Curated, safe-for-work showcase — every image ships with its prompt and settings.
How to run Chroma
Two paths — one takes ten seconds, one takes an afternoon.
⚡ Run on Civitai (Recommended)
The fastest way to start.
🖥️ Run it locally
For power users who want full control.
No graphics card? The Civitai path above skips all of this.
Chroma vs other ecosystems
Backed by Civitai usage data.
| Feature | Chroma | Flux | SDXL | Qwen |
|---|---|---|---|---|
| Best for | Neutral open base for fine-tuning | Polished photorealism & text | Versatile realism & huge LoRA depth | Prompt accuracy & in-image text |
| Prompt style | Natural language | Natural language | Tags | Natural language |
| License | Apache 2.0 | Non-commercial (Dev) | CreativeML / OpenRAIL | Apache 2.0 |
| Parameters | ~8.9B | 12B | ~3.5B | 20B |
| LoRA ecosystem | Growing | 40,000+ | 40,000+ | 1,600+ |
| Available on Civitai | ✓ Yes | ✓ Yes | ✓ Yes | ✓ Yes |
Frequently asked questions
How much does it cost to generate with Chroma?
What exactly is Chroma?
How is Chroma different from Flux?
Can I train my own Chroma LoRA?
Which Chroma version does Civitai host?
Do I need a GPU to run Chroma?
Start generating with Chroma now
No installation. No GPU. Runs in your browser.