How to run Chroma locally.

What Chroma is.

Chroma is a community model by lodestones, funded by donations. It started from Flux.1 Schnell and was retrained into an 8.9B model: a 3.3B layer was swapped for a much smaller one, and the T5 padding is masked out. It uses T5-XXL only, with no CLIP-L, and the Flux.1 VAE.

• Chroma1-HD is the one to use, uploaded in August 2025.
• Chroma1-Base is the base checkpoint, published shortly before HD.
• Chroma1-Flash is a faster version with the guidance baked in. lodestones points to its delta weights for low-step use.
• Chroma1-Radiance is a different, pixel-space model with no VAE and its own nodes. It’s still in progress and not covered here.

Unlike Flux.1 Dev and Schnell, Chroma uses a real CFG, so the negative prompt works. It prefers natural sentences to tag lists, in the negative prompt too.

Files you need.

If you already run Flux.1, you have the T5 and the VAE. Chroma loads its T5 through a single Load CLIP node with type chroma.

• Model Chroma1-HD-fp8mixed.safetensors ComfyUI/models/diffusion_models/ full precision: Chroma1-HD.safetensors, 17.8 GB
  9.2 GB Download
• Text encoder t5xxl_fp8_e4m3fn_scaled.safetensors ComfyUI/models/text_encoders/ or t5xxl_fp16.safetensors, 9.8 GB
  5.2 GB Download
• VAE ae.safetensors ComfyUI/models/vae/
  0.3 GB Download

The full-precision model is at lodestones/Chroma1-HD. Users find a plain fp8 conversion noticeably worse than full precision; Comfy’s fp8mixed file, a Q8 GGUF or the full file are the ones to use.

models/
├── diffusion_models/
│   └── Chroma1-HD-fp8mixed.safetensors
├── text_encoders/
│   └── t5xxl_fp8_e4m3fn_scaled.safetensors
└── vae/
    └── ae.safetensors

Smaller files: GGUF

silveroxides’ GGUF builds load through the ComfyUI-GGUF nodes and go in diffusion_models.

Q8_0 is close to the original. Q6_K and Q5_K_M are the usual middle ground.

What fits your computer.

Chroma1-HD at 1024 × 1024 with the fp8 T5. ComfyUI runs the encoder first and moves it aside, so the model file sets the limit.

• 6 GBTight

  Q4_K_M GGUF (5.6 GB), with the T5 in system RAM.

• 8 GBOffloads

  Q5_K_M or Q6_K GGUF. Users report it working on 8 GB with offloading.

• 12 GBFits

  The fp8mixed model or Q8_0 GGUF. One 12 GB user ran the full-precision model too.

• 16 GBFits

  fp8mixed or Q8_0 comfortably.

• 24 GBFits

  Full precision (17.8 GB), with the T5 moved aside during sampling.

• MacUntested

  Use the full-precision file or a GGUF: a Mac can’t compute in fp8. We found no Mac reports for Chroma.

• 8 GB graphics cards
• 12 GB cards
• 16 GB cards
• 24 GB cards
• Mac with Apple Silicon

Set it up.

1. 1

   Download the three files

   Model, T5 and VAE from the list above. None of them is gated.

2. 2

   Put them in their folders

   Model in diffusion_models, T5 in text_encoders, VAE in vae. Then restart ComfyUI so the loaders list them.

3. 3

   Open the Chroma template

   In ComfyUI’s template browser, pick Chroma: Text to Image. lodestones also publishes a ComfyUI workflow of their own on the model page.

4. 4

   Check the loaders

   Load Diffusion Model gets the Chroma file. Load CLIP gets the T5 with type chroma. Load VAE gets ae.safetensors. For a GGUF, swap the model loader for Unet Loader (GGUF).

5. 5

   Leave the shift and padding alone

   ModelSamplingAuraFlow at shift 1 is the value the creator intended. T5TokenizerOptions sets the T5 padding; Comfy’s note says a min_padding of 1 is the official way and 0 works too.

6. 6

   Write both prompts and run

   Describe the image in sentences, and write the negative prompt in sentences too, like the template’s “This low quality greyscale unfinished sketch is inaccurate and flawed.”

Settings that work.

These are from lodestones’ own ComfyUI workflow for Chroma1-HD.

Steps

26

CFG

3.8

Sampler

euler

Scheduler

BetaSamplingScheduler

Alpha, beta

0.45, 0.45

Shift

1

Size

1152 × 1152

Negative

yes, in sentences

Comfy’s Chroma template is close: 26 steps, CFG 3.5, the plain beta scheduler and 1024 × 1024. The diffusers example on the model card uses 40 steps at guidance 3.0. Keep CFG between about 3 and 5; higher burns the image. lodestones says any ODE sampler should work, with beta or sigmoid schedules for fewer steps. Stay under about 2 megapixels, where quality starts to fall off.

How fast.

User reports. The 1.3 s figure used torch.compile and SageAttention as well.

When it goes wrong.

Messy or burnt images

CFG is too high. Stay around 3 to 5. For more contrast without burning, users try CFGNorm or Skimmed CFG.

Soft, low-detail images from an fp8 file

A plain fp8 conversion loses more than you’d expect. Switch to Comfy’s fp8mixed file, a Q8_0 GGUF or full precision.

The prompt is only half followed

Chroma reads sentences better than tags. Rewrite tag lists as a description, and do the same in the negative prompt.

Perturbed Attention Guidance does nothing

PAG only works on UNet models like SDXL. Chroma is a transformer, so the node has no effect.

Faces and details fall apart at large sizes

Above about 2 megapixels, Chroma degrades. Generate near 1 megapixel and upscale.

Questions.

Sources: Chroma1-HD model card, lodestones’ ComfyUI workflow, Comfy-Org Chroma1-HD files, fp8 vs bf16 thread [1], RTX 5090 thread [2], CFG thread, sampler thread, ComfyUI issue #12177.