How to run Chroma text-to-image in ComfyUI on a cloud GPU
Rent a GPU on Deploy, then open a terminal in JupyterLab or the ComfyUI page. Install ComfyUI from the official repository, place the files the Chroma card and the official ComfyUI template name, load the template workflow, and copy the output folder off the instance before you destroy it. Destroying the instance deletes its disk.
The ComfyUI install, whose README’s memory clause is 4GB VRAM, is How to install ComfyUI and ComfyUI-Manager on a cloud GPU. FLUX.1-dev file placement in ComfyUI is How to run FLUX.1-dev in ComfyUI on a cloud GPU. HiDream-I1, whose README publishes no VRAM figure, is How to run HiDream-I1 on a cloud GPU. Qwen-Image, whose README and model cards publish no diffusers VRAM figure, is How to run Qwen-Image on a cloud GPU. Sana, whose installation page lists 9GB, 12GB, and less than 8GB, is How to run Sana text-to-image on a cloud GPU. How published FLUX and SDXL weight sizes are read: How much VRAM do I need for FLUX and SDXL. If the process runs out of memory: How to fix CUDA out of memory on a cloud GPU. ComfyUI listens on 127.0.0.1 port 8188: How to use SSH port forwarding on a cloud GPU. FLUX.1 Krea [dev], whose ComfyUI tutorial prints greater than 32GB and like 24GB, is How to run FLUX.1 Krea [dev] in ComfyUI on a cloud GPU.
The lodestones/Chroma card says “THIS REPO IS DEPRECATED!” and says to use Chroma1-HD, Chroma1-Base, or Chroma1-Flash instead. The same card says Chroma is an 8.9B parameter model based on FLUX.1-schnell and that it is Apache 2.0 licensed. cardData.license on 2026-11-13 was apache-2.0, gated was false, and the tags included not-for-all-audiences. lastModified was 2025-10-23T09:09:18.000Z. The model APIs for lodestones/Chroma1-HD, lodestones/Chroma1-Base, and lodestones/Chroma1-Flash that day also reported cardData.license apache-2.0 and gated false. The ComfyUI knowledge page site/knowledge/models/chroma.md says Chroma is Apache 2.0 licensed and uses the words “unrestricted use” next to that. Read the model card before you redistribute the weights.
The steps below are copied from the files named in each section. They were not run on a CloudGPU instance for this page. Anything this page adds around them is marked.
What the docs say about memory
The lodestones/Chroma card’s how-to section prints no GPU GB figure. site/knowledge/models/chroma.md in Comfy-Org/workflow_templates says the base model is 24GB VRAM recommended (BF16), Q8_0 quantized is ~13GB VRAM, and Q4_0 quantized is ~7GB VRAM. It also says the model requires a FLUX.1 VAE and a T5 text encoder. Those sentences are that file’s. This page did not remeasure them. The same file describes Chroma XL as an experimental merge, with low CFG (2.5-3.0) and a low step count (8-12 steps). The text-to-image template below does not load a file named Chroma XL.
The official template note prints parenthetical sizes next to the download links: 16.58 GB, 8.56 GB, 9.12 GB, 4.8 GB, and 319.8 MB. Those are the note’s file sizes. This page does not treat them as VRAM. The note’s fp8 line says “fp8_scaled version for lower VRAM GPU” and does not print a GB figure for that phrase.
GET /v1/public/compute-capabilities does not list GPUs. It returns feature flags. The memory column is vram_gb from gpu-prices.json, updated_at 2026-11-13T20:10:32.168Z. GET /v1/public/gpu-catalog the same day (generated_at 2026-11-13T20:10:33.209Z) listed the same names and the same vram_gb values, plus Ascend 910B at 64, which is not in gpu-prices.json and is not listed here. The catalog response also includes an availability field. This page does not copy it. Every prices row has gpu_count 1. Every catalog row has min_gpu_count 1. This page does not mark a row as enough. The knowledge page’s numbers are 24, about 13, and about 7. The catalog has rows at 24 and no row at 13 or 7. Which listing you can start is on Deploy.
The catalog template list had no Chroma template.
| Catalog name | vram_gb |
|---|---|
| RTX 2080 | 8 |
| RTX 3080 | 10 |
| RTX 2080 Ti | 11 |
| RTX 3080 Ti | 12 |
| V100 16G | 16 |
| A30 | 24 |
| RTX 3090 | 24 |
| RTX 4090 | 24 |
| RTX 4090D | 24 |
| RTX 4080 SUPER 32G | 32 |
| RTX 5090 | 32 |
| RTX 5090D | 32 |
| V100 32G | 32 |
| A100 40G | 40 |
| A800 40G | 40 |
| A40 | 48 |
| RTX 4090 48G | 48 |
| RTX 4090D 48G | 48 |
| RTX A6000 | 48 |
| NVIDIA RTX PRO 5000 72G | 72 |
| RTX PRO 5000 72G | 72 |
| A100 80G | 80 |
| A800 80G | 80 |
| H800 | 80 |
| RTX 6000D | 83 |
| H20 96G | 96 |
| RTX PRO 6000 96G | 96 |
| RTX PRO 6000 Max-Q | 96 |
| H20-3e 141G | 141 |
Files and folders
Install ComfyUI with How to install ComfyUI and ComfyUI-Manager on a cloud GPU. The card’s how-to, under the heading ComfyUI, says:
- put
T5_xxlintoComfyUI/models/clip - put
FLUX VAEintoComfyUI/models/vae - put
Chroma checkpointintoComfyUI/models/diffusion_models - load the chroma workflow
- run the workflow
The same section links the T5 files t5xxl_fp16.safetensors and t5xxl_fp8_e4m3fn.safetensors on comfyanonymous/flux_text_encoders, the VAE ae.safetensors on lodestones/Chroma, and the workflow ComfyUI_Chroma1-HD_T2I-workflow.json on lodestones/Chroma. It says to pick the latest checkpoint on that repo. A section titled “Deprecated: Manual Installation (Chroma)” clones https://github.com/lodestone-rock/ComfyUI_FluxMod.git into ComfyUI/custom_nodes. That heading says deprecated. This page does not use that clone as the current route.
The official template is templates/image_chroma_text_to_image.json in Comfy-Org/workflow_templates. Its MarkdownNote says the final model is hosted at lodestones/Chroma1-HD and that previous versions are at lodestones/Chroma. The note’s tree, copied, is:
ComfyUI/
└─ models/
├─ diffusion_models/
│ ├── Chroma1-HD-fp8_scaled_rev2_learned_topp0.25_k16-2688_svd.safetensors
│ └── Chroma1-HD.safetensors
├─ text_encoders/
│ ├── t5xxl_fp8_e4m3fn_scaled.safetensors
│ └── t5xxl_fp16.safetensors
└─ vae/
└── ae.safetensors
That tree puts the text encoder in models/text_encoders. The card’s step 1 says ComfyUI/models/clip. Both sentences are upstream. This page does not merge them into one path. The note’s download lines, with the file sizes labelled as the note prints them, are Chroma1-HD.safetensors (16.58 GB) on lodestones/Chroma1-HD, Chroma1-HD-fp8mixed.safetensors (8.56 GB) on Comfy-Org/Chroma1-HD_repackaged under split_files/diffusion_models/, t5xxl_fp16.safetensors (9.12 GB) and t5xxl_fp8_e4m3fn_scaled.safetensors (4.8 GB) on comfyanonymous/flux_text_encoders, and ae.safetensors (319.8 MB) on Comfy-Org/Lumina_Image_2.0_Repackaged under split_files/vae/. The note also lists ModelScope mirrors. This page copies the Hugging Face lines.
The template’s loader widgets are UNETLoader Chroma1-HD-fp8mixed.safetensors with weight dtype default, CLIPLoader t5xxl_fp8_e4m3fn_scaled.safetensors with type chroma and type default, and VAELoader ae.safetensors. SaveImage’s widget is the prefix ComfyUI. A note in the template says the flow shift node set to 1.0 represents the shift the creator of the Chroma model intended. Another note says min_padding 1 is the official way and that setting it to 0 works too. The model API for Comfy-Org/Chroma1-HD_repackaged on 2026-11-13 listed split_files/diffusion_models/Chroma1-HD-fp8mixed.safetensors and split_files/diffusion_models/Chroma1-HD.safetensors, with cardData.license apache-2.0 and gated false. This page did not download the files and did not run the workflow.
Web UI
The card and the template do not document a Gradio launch. ComfyUI is the UI. How to install ComfyUI and ComfyUI-Manager on a cloud GPU says comfy/cli_args.py defaults --listen to 127.0.0.1 and --port to 8188. How to use SSH port forwarding on a cloud GPU is the tunnel. The docs.comfy.org sitemap on 2026-11-13 had built-in node pages ChromaRadianceOptions and EmptyChromaRadianceLatentImage. Those pages describe Chroma Radiance options, including a note that non-tiling mode may require a lot of VRAM, and they do not print a GPU GB figure. They do not document the lodestones/Chroma folder steps. This page does not treat them as the text-to-image template.
When it fails
CUDA out of memory. The card’s how-to prints no GB cutoff. The knowledge page’s figures are 24GB recommended for the BF16 base, ~13GB for Q8_0, and ~7GB for Q4_0. The template’s fp8 line does not print a GB figure. If the process raises that error, use how to fix CUDA out of memory on a cloud GPU.
“CUDA error: no kernel image is available for execution on the device” on an RTX 5090. The Chroma card and the template do not pin a PyTorch build. The ComfyUI install guide’s README line for current NVIDIA cards is a separate page. The RTX 5090 guide says that card is compute capability 12.0 (sm_120) and that PyTorch 2.7 and later, on a build for CUDA 12.8 or newer, is the line that announced support. This page did not install a torch wheel and did not re-test this workflow on an RTX 5090. Catalog RTX 5090 and RTX 5090D are 32. That is the catalog number only.
Save the images, then destroy
The template’s SaveImage prefix is ComfyUI. The install guide says folder_paths.py sets output_directory to output, and that --output-directory overrides it. Copy that directory before you destroy the instance. Save your work before you destroy a cloud GPU is the checklist. The terms say data on local instance storage is permanently deleted when the instance is destroyed.
The pricing page FAQ states the billing rule in these words:
Per-minute options have a 1-minute minimum; unused prepaid minutes are refunded when you destroy the instance. Some configurations use whole-hour billing with a 1-hour minimum; unused time within a charged hour is not refunded on voluntary stop. The billing period, minimum charge and initial prepayment are shown before confirmation. No subscription.
Prices are not copied here.
Frequently asked questions
How much VRAM do the Chroma docs print?
The lodestones/Chroma card how-to prints no GPU GB figure. site/knowledge/models/chroma.md says 24GB VRAM recommended for the BF16 base, ~13GB for Q8_0, and ~7GB for Q4_0. The template note parenthetical numbers are file sizes, and the fp8 line does not print a GB figure. This page does not mark a catalog row as enough. Which listing you can start is on https://cloudgpu.app/deploy.
Which files go in which ComfyUI folders?
The card says T5_xxl goes in ComfyUI/models/clip, the FLUX VAE in ComfyUI/models/vae, and the Chroma checkpoint in ComfyUI/models/diffusion_models. The official template tree puts Chroma1-HD.safetensors in models/diffusion_models, the T5 files in models/text_encoders, and ae.safetensors in models/vae. The template loaders name Chroma1-HD-fp8mixed.safetensors, t5xxl_fp8_e4m3fn_scaled.safetensors with type chroma, and ae.safetensors. The card says the lodestones/Chroma repo is deprecated and names Chroma1-HD, Chroma1-Base, and Chroma1-Flash. This page did not download the files.
Where are the images saved?
The template SaveImage widget is the prefix ComfyUI. The ComfyUI install guide says folder_paths.py sets the output directory to output. The card and the template do not document a Gradio port. The default ComfyUI listen address in the install guide is 127.0.0.1 port 8188. This page did not run the workflow.
What license does the model card print?
The lodestones/Chroma card says the model is Apache 2.0 licensed. The model API on 2026-11-13 reported cardData.license apache-2.0, gated false, and a not-for-all-audiences tag. Chroma1-HD, Chroma1-Base, and Chroma1-Flash reported the same license field and gated false. The ComfyUI knowledge page says Apache 2.0 and uses the words unrestricted use. That wording is license text, not a CloudGPU price and not a storage allowance.
Sources (read 2026-11-13, Asia/Shanghai)
- Hugging Face model card
lodestones/Chroma(deprecation line, 8.9B parameter sentence, Apache 2.0, ComfyUI folder steps, deprecated ComfyUI_FluxMod clone,not-for-all-audiencestag): read 2026-11-13. Model API the same day:gatedfalse,cardData.licenseapache-2.0,lastModified2025-10-23T09:09:18.000Z,sha2f3b2730d7b5edbd02cdff12f72610af5787300b. The support paragraph that names a card this page does not print is not copied - Model APIs
lodestones/Chroma1-HD,lodestones/Chroma1-Base,lodestones/Chroma1-Flash, andComfy-Org/Chroma1-HD_repackaged:gatedfalse,cardData.licenseapache-2.0, read 2026-11-13. Chroma1-HDlastModified2025-10-23T09:07:42.000Z. The repackaged file list includedsplit_files/diffusion_models/Chroma1-HD-fp8mixed.safetensorsandsplit_files/diffusion_models/Chroma1-HD.safetensors - Comfy-Org/workflow_templates
site/knowledge/models/chroma.md(24GB, ~13GB, ~7GB): commita2bcdabeb3c198951d92f8c307ce1d73c9e42393, committed 2026-02-08T03:07:14Z - Comfy-Org/workflow_templates
templates/image_chroma_text_to_image.json(folder tree, loader widgets, SaveImage prefixComfyUI, file-size parentheticals): commitea7faae6999fb31338f43408ff4a575bd3f7d9a0, committed 2026-11-08T16:45:29Z - docs.comfy.org built-in nodes
ChromaRadianceOptionsandEmptyChromaRadianceLatentImage: read 2026-11-13. No GPU GB figure. Not used as the text-to-image folder steps - Public GPU memory: https://cloudgpu.app/gpu-prices.json (
updated_at2026-11-13T20:10:32.168Z) and https://cloudgpu.app/v1/public/gpu-catalog (generated_at2026-11-13T20:10:33.209Z).GET /v1/public/compute-capabilitiesreturned feature flags and no GPU list. The catalog had no Chroma template
Cards you can start are on the Deploy page. Prices are on https://cloudgpu.app/pricing and are not repeated here.
Last updated: 2026-11-13 (Asia/Shanghai).