Render on RunPod
No big graphics card, or one that is busy? bitGen Studio can send its renders to a GPU you rent from RunPod and bring the pictures, videos and songs back. Here is every step - and what changes when the GPU is not yours.
Make a RunPod account bitGen's referral link: signing up with it helps fund bitGen Studio.
- Read this first
- Serverless or a pod?
- Your RunPod account
- Put your models on RunPod
- Make the Serverless endpoint
- Connect bitGen Studio
- A first test with SDXL
- Which styles work
- Only some work on RunPod
- Using a pod instead
- When something goes wrong
Read this first
Your own models are not used. RunPod's GPUs are in RunPod's data centres and cannot read anything on your computer. Every model, LoRA and VAE a render uses must be on RunPod first - copied once to a network volume there, where it stays for every render after.
- Custom node packs must be on RunPod too: they go into the worker's image. Many styles need none - see Which styles work.
- Serverless: a render and its input files (reference pictures, videos, songs) must be under 10 MB together - RunPod's limit. bitGen checks before it sends anything.
- Serverless: the first render after a quiet spell waits for a worker to start - a cold start can take minutes. There are no live steps while it runs, only the time.
- bitGen's check for missing models before a render is skipped - it cannot see RunPod's disk. A model that is not there fails the render with RunPod's own message.
- Your prompts, reference pictures and results travel to RunPod and back over the internet. Rendering on your own computer keeps them at home.
- It costs money: RunPod charges for the GPU while it runs and a monthly fee for the network volume. Prices change - see runpod.io/pricing. bitGen adds what you tell it a GPU costs per hour to 📊 Usage.
Serverless or a pod?
| Serverless endpoint | Pay only while a render runs. No machine to start or stop. Best for a render now and then. Limits: 10 MB a job, cold starts, no live steps. |
|---|---|
| Pod | A whole GPU machine running ComfyUI, paid by the hour while it is on - you start and stop it. Best for a long session of renders. No 10 MB limit, live steps, and node packs can be added from ComfyUI. |
Both need your models on RunPod. The steps below set up Serverless; a pod is under Using a pod instead.
Your RunPod account
- Sign up and add creditat runpod.io - renders are paid from your balance.
- Make an API keySettings ▸ API Keys ▸ Create API Key. bitGen keeps it in its settings on your computer. Keep it to yourself: anyone with it can spend your balance.
Put your models on RunPod
A network volume is a disk in one of RunPod's data centres that keeps its files. Serverless workers see it as /runpod-volume; a pod sees the same disk as /workspace.
- Make a network volumeStorage ▸ New Network Volume. Pick a data centre that has the GPUs you want to render on - the endpoint can only use GPUs in the volume's data centre. Make it big enough for your models (a video model with its text encoder can be 30 GB or more).
- Start a pod to fill itPods ▸ Deploy. Choose the network volume first (only GPUs in its data centre are then shown) - it can only be attached when the pod is made. Take the cheapest GPU and the RunPod PyTorch template: it has a web terminal.
- Download the models straight onto the volumeConnect ▸ Jupyter, open a terminal and download each model into the folder ComfyUI reads it from, under /workspace/models. A pod downloads at data-centre speed - much faster than sending files from home. For example, ACE-Step 1.5:
The folders are ComfyUI's own: checkpoints, diffusion_models, loras, text_encoders, vae and so on. bitGen's 🧩 Model manager shows where each of your models came from.cd /workspace && mkdir -p models/diffusion_models models/text_encoders models/vae HF=https://huggingface.co/Comfy-Org/ace_step_1.5_ComfyUI_files/resolve/main/split_files wget -c -P models/diffusion_models $HF/diffusion_models/acestep_v1.5_turbo.safetensors wget -c -P models/text_encoders $HF/text_encoders/qwen_0.6b_ace15.safetensors wget -c -P models/text_encoders $HF/text_encoders/qwen_1.7b_ace15.safetensors wget -c -P models/vae $HF/vae/ace_1.5_vae.safetensors - Models only you haveA LoRA you trained or a model you merged is not online to download - send it from your computer: drag it into the pod's file manager (RunPod's file browser template), or use RunPod's S3-compatible API for network volumes. Your upload speed decides how long it takes.
- Stop the podTerminate it when the files are in - it charges while it runs. The volume keeps the files.
The model's file name is how a style finds it, folders below models/checkpoints included - a file at models/checkpoints/sdxl/juggernaut.safetensors is sdxl/juggernaut.safetensors. In bitGen, write that name in the style's Model box: the list there shows the models on your computer, not on RunPod.
Make the Serverless endpoint
The endpoint runs RunPod's own ComfyUI worker (worker-comfyui, version 5) - bitGen talks to it over the internet.
- New endpointServerless ▸ New Endpoint ▸ Deploy from a Docker image. (Not "Deploy image or video API" or the Hub - they make a different kind of worker.)
- The image
runpod/worker-comfyui:5.10.0-base- ComfyUI with no models; yours come from the volume. For node packs your styles need, build your own image from it (see Which styles work). - The settings
GPU enough memory for your biggest model - 16 GB for SDXL, 24-48 GB or more for video models; tick a few types so a worker is easier to get Active workers 0 - nothing runs, or costs, while you are not rendering Max workers 1 to start with - it caps the spend Idle timeout 5 seconds; FlashBoot on Container disk 20 GB Network volume the one with your models Start command, variables leave empty - the image starts the worker itself - Deploy, and copy the endpoint IDIt is on the endpoint's page, and in its address:
https://api.runpod.ai/v2/abc123xyz/run- the ID is the part after/v2/.
RunPod's page shows a curl example with a prompt - bitGen does not need it: it sends the whole ComfyUI workflow itself.
Connect bitGen Studio
- Settings ▸ Remote generation ▸ RunPod ServerlessPaste the API key and the endpoint ID. Add what a worker costs per hour if you want it in 📊 Usage (RunPod shows it per second - × 3600).
- ✓ TestRunPod answers with its workers and jobs. "0 ready" is normal: the first render starts a worker.
- Use itEvery render now goes to the endpoint - pictures, videos and songs alike. This computer switches back.
On the way, bitGen adapts each render to the worker, which only sends back pictures and files saved by ComfyUI's Save Video node: a video saved with Video Combine is saved with Save Video instead, and a song travels inside a tiny silent video and is taken out again on your computer. You get the same files as from your own GPU.
A first test with SDXL
To see it work before moving any models, make the endpoint from runpod/worker-comfyui:5.10.0-sdxl with no network volume: that image has SDXL built in (sd_xl_base_1.0.safetensors), the model bitGen's SDXL style uses. Then pick the SDXL style, leave its Model box empty, remove any LoRAs, and render. The first picture can take several minutes while RunPod starts a worker; the next comes in seconds.
Which styles work
From version 2.749.0 bitGen comes with ready-made remote styles for RunPod's own images - nothing to upload: R_SDXL and R_SDXL img2img (the sdxl image), R_FLUX.1 schnell, R_FLUX.1 dev and R_FLUX.1 dev fp8 (the flux1-schnell, flux1-dev and flux1-dev-fp8 images), R_SD3 Medium (sd3) and R_Z-Image Turbo (z-image-turbo). Each image holds one set of models, so bitGen asks RunPod which image your endpoint runs and lists only its styles (Settings ▸ Remote generation ▸ Your endpoint's image).
RunPod's worker image carries ComfyUI 0.34.0 and no node packs. Checked against it:
| Work as they are | MiniMax H3 video (reference pictures, first and last frame, the scene board), ACE-Step 1.5 and MiniMax Music3 songs, and most image styles: SDXL, FLUX 2 and FLUX 3 Klein, FLUX Kontext and Fill, Krea 2, Qwen Image Edit 2511, Anima, Moss. |
|---|---|
| Need a newer ComfyUI | YuE2, Fibo 2, Qwen Image 2.1, PH3 and FastH3 Omni use nodes added to ComfyUI after 0.34.0 - they need your own image built with a newer ComfyUI (the worker's build has a COMFYUI_VERSION setting), or a pod. |
| Need node packs | The GGUF styles (ComfyUI-GGUF), character swap (VideoHelperSuite, rgthree), LTX Ripple and the LTX canvas (KJNodes, LTXVideo, VideoHelperSuite, RES4LYF), Viggle, the character sheet creator (H3LookSheets, KJNodes) and the music video with a song (H3 NativeAudioLock). |
While RunPod Serverless is in use, bitGen's style lists show only the styles the endpoint can render (from version 2.744.0). In Settings ▸ Remote generation, say whether the image has a newer ComfyUI or extra node packs, and see the hidden styles with the reason.
Your own image starts from the worker's, adds the packs and goes to RunPod from GitHub (Deploy from a GitHub repository) or Docker Hub - the worker's customization guide shows how. A pod is the easier way for these.
FROM runpod/worker-comfyui:5.10.0-base
RUN comfy-node-install comfyui-videohelpersuite comfyui-kjnodes
Only some work on RunPod
Basic Settings ▸ GPU decides what runs where (from version 2.745.0): Remote sends everything to RunPod, Local keeps everything on your computer, and Both lets you tick what goes to RunPod - the LLM, the vision LLM, pictures, and video + songs. With Both, your own ComfyUI starts as usual and each ticked job is sent to RunPod when it starts.
The LLM and the vision LLM run on a second Serverless endpoint made with RunPod's vLLM worker (Serverless ▸ New Endpoint ▸ Deploy LLM API, with the Hugging Face model it should run). Put its ID in Settings ▸ Remote generation ▸ LLM on RunPod. The vision LLM needs a model that takes pictures, such as Qwen2.5-VL - one such endpoint can do both.
From version 2.746.0 each style says where it runs: in its settings (Settings ▸ Image, Video, or Audio ▸ Music styles) set Runs on to Remote GPU, and its name starts with R_. Local lists only the styles for your computer, Remote only the R_ styles, and Both lists them all - an R_ style renders on RunPod, the others on your computer.
A remote style can have its own endpoint ID, so one endpoint can hold the SDXL models and another the FLUX ones. Leave it blank to use the endpoint in Settings ▸ Remote generation.
Using a pod instead
- Deploy a ComfyUI podPods ▸ Deploy with your network volume and a ComfyUI template that exposes ComfyUI's port (8188) over HTTP. Add node packs from ComfyUI's Manager in the pod.
- Settings ▸ Remote generation ▸ RunPod podPaste the API key and press 🔄 My pods: your pods with their GPU, price and state. Pick one, ▶ start it, set the port, ✓ Test, Use it.
- ⏹ Stop it when you are doneA pod charges by the hour while it is on, rendering or not.
The models are the same story: on the pod's volume, not your computer. Keep a pod's ComfyUI off the open internet when you can - RunPod's proxy address is hard to guess, but anyone who has it can use your GPU.
When something goes wrong
| The render ran on my own computer | Settings ▸ Remote generation must say "now: RunPod Serverless" - press Use it there. Before version 2.744.0, pressing Save in Settings switched back to this computer; update bitGen. |
|---|---|
| A style is missing from the list | Basic Settings ▸ GPU hides the styles that do not run there: on Local the R_ styles, on Remote the others. Change the style's Runs on, or pick Both. |
| "RunPod refused the API key" | Copy the key again, or make a new one (Settings ▸ API Keys). |
| "RunPod does not know the endpoint" | The endpoint ID is only the part after /v2/ - no slashes, no /run. |
| The render waits a long time | A cold start, or no GPU of the chosen type free in that data centre - tick more GPU types on the endpoint. Its page shows the workers starting. |
| "Value not in list" for a model | The model is not on the volume, its name is spelled differently, or the volume is not attached to the endpoint. |
| A node "does not exist" | The style needs a node pack or a newer ComfyUI than the image has - see Which styles work. |
| "RunPod takes 10 MB at most" | Use shorter or smaller reference files, or render this one on a pod or your own computer. |
| "sent no files back" | The endpoint runs an old worker - use worker-comfyui 5 or later. |
The endpoint's Logs tab on RunPod shows the worker's side. Still stuck? Ask in the Discussions.