Remote generation
Render bitGen Studio's pictures, videos and songs on a bigger graphics card somewhere else - a GPU you rent from RunPod, or another computer of yours - and get them back in the studio.
Using bitGen 2.754.0 or older? RunPod switches off the API it used for pods on 15 November 2026. Update to 2.755.0 or newer (Settings ▸ General ▸ Updates, or git pull in the bitGen folder) so My pods, start and stop keep working.
- How it works
- Four kinds of remote GPU
- RunPod: tested
- RunPod pods: the PODS docker
- OpenRouter: the LLM
- Online services: Settings ▸ API
- Ready-made remote styles
- Endpoints for all your styles
- Costs and limits
How it works
- Settings ▸ Remote generation ▸ Use remote generation turns it on. Off, everything runs on your computer and Basic Settings ▸ GPU is hidden.
- With Local picked, Basic Settings ▸ GPU shows nothing remote - no Remote servers or RunPod choices, no Remote generation, PODS or 💲 RunPod buttons. Pick Remote or Both to see them again.
- Settings ▸ Remote generation is where you set up the GPU: RunPod, another computer or an SSH tunnel.
- Basic Settings ▸ GPU ▸ Remote servers picks RunPod - Serverless (an endpoint, paid only while a job runs) or Pods (your own GPU machine, paid by the hour while it runs) - or OpenRouter (the LLM only; renders stay on your computer).
- Basic Settings ▸ GPU decides what goes there: Local (the default - everything on your computer), Remote (everything on the remote GPU) or Both (each style decides; jobs with no style follow the Image and Video + songs boxes).
- Each style has Runs on: This computer or Remote GPU. A remote style's name starts with R_, and it can have its own RunPod endpoint.
- The job's log says where it rendered, and 📊 Usage adds what the remote GPU cost when you give its price per hour.
Four kinds of remote GPU
| RunPod Serverless | An endpoint you pay for only while a render runs. It runs RunPod's ComfyUI worker; the models are in its image or on a network volume. |
|---|---|
| RunPod pod | A whole GPU machine running ComfyUI, paid by the hour while it is on. The studio lists your pods and starts or stops them. |
| Another computer | A ComfyUI on another computer on your network - give its address, for example http://192.168.1.20:8188. |
| SSH tunnel | A ComfyUI on a server you reach over SSH. The studio opens the tunnel, so ComfyUI never faces the internet. |
RunPod: tested
Make a RunPod account bitGen's referral link: signing up with it helps fund bitGen Studio.
Confirmed working: bitGen Studio rendered SDXL pictures on a RunPod Serverless endpoint running RunPod's own runpod/worker-comfyui:5.10.0-sdxl image. The picture came back in under a minute once a worker was running, and RunPod's log showed the same seed as the studio's.
Video, songs and the other image styles are checked against the ComfyUI in RunPod's worker, but not yet tried on a real endpoint. Your reports help - tell us in the Discussions what worked.
Every step - the account, the API key, your models, the endpoint, connecting the studio and what to do when something goes wrong - is on Render on RunPod.
RunPod pods: the PODS docker
From version 2.756.0 you can find, create and run RunPod pods without leaving bitGen. A pod is your own GPU machine running ComfyUI; you pay for it by the hour while it runs.
Allow time for the first start. Creating a new pod and setting it up takes about 10-15 minutes, plus the time to transfer your models to it. Files the pod downloads itself (Hugging Face, Civitai) arrive quickly; files sent from your computer go at your upload speed. After that the pod starts in a few minutes.
- Pick PodsBasic Settings ▸ GPU ▸ Remote servers: RunPod, then Pods. The 🖥 PODS button appears there.
- Create a bitGen podFrom version 2.780.0 there is no template to pick: + Create pod in the PODS docker makes a bitGen pod. bitGen picks RunPod's PyTorch template and installs ComfyUI, the custom nodes your styles use and their files on it - a few minutes the first time, then kept in /workspace.
- Create the podIn the form pick the GPU (RunPod's price per hour now, its memory and how many are free), Secure or Community cloud (usually cheaper), the container disk and /workspace's size. The hourly estimate shows before you create it.
- Use itA bar shows how far it has got - waiting for a GPU machine, starting the container, ComfyUI starting (a template may download its models now), then "pod loaded". The time it took is kept, so the next start shows "3 min of about 7 min". Press Use it to render there.
- Your styles on the podIn My pods, open a pod and pick your styles in Styles (the Generate mode's, or All), then press 💾 Save: bitGen puts each style's workflow, models, LoRAs and node packs on the pod. Tick Image, Image edit or Video on a pod to send every style of that kind there - a pod for each kind, or all on one pod. With Pods picked your own styles show, and P_ styles made before still work.
A style with no pod yet: press Generate and the PODS docker opens with the pod already filled in - a bitGen pod with the style and its kind, the cheapest GPU with enough memory (its price an hour) and the disk. Pick the GPU and how many, then Create pod - or press ▶ Use for the style on one of your pods. The style remembers its pod: if the pod is stopped bitGen starts it, and if it is gone bitGen makes a new one the same way. Files in bitGen's download list are downloaded by the pod itself; other files are sent from your computer. Basic Settings shows what your pods cost (RunPod's own bill, every minute) and 📊 Usage keeps it by day.
Getting models onto a pod: On a running pod, ⬇ Models downloads a style's files - or single files from bitGen's download list - into its ComfyUI, through the pod's Jupyter (bitGen sets its own password on pods it creates; for others, paste it) or ComfyUI-Manager (only the models in its own list).
Only the custom nodes your workflows use go with it, cloned on the pod from their git address: a pod gets the custom node packs its styles' workflows use and it lacks - found in bitGen's node list, the pack's own pyproject.toml, ComfyUI-Manager's list or the Comfy Registry, at the same version - then its ComfyUI restarts. Only a pack with no address (one you made) or a private one is uploaded from your computer. ⬆ Nodes its styles need in a pod's item does the same for any pod. While a render waits for its pod, a line over the status bar and the Queue say what the pod is doing - a model being sent shows how far it is.
In the PODS docker each running pod shows how busy its GPU is and how full its GPU memory, RAM and disks are. Before a pod is made, the form lists everything it will be - the template, the GPUs, the price an hour, the disks and what goes on it - and a pod RunPod no longer has is never made again without asking you.
The 💲 RunPod docker shows your RunPod balance, what it is spending now and how long the balance lasts at that, what your pods cost today and this month, and RunPod's GPU prices on the Secure and Community clouds - open it from Basic Settings ▸ GPU or the PODS docker. It goes inside your own dockers like PODS and Basic Settings do, and the floating dialogs stay in front of every docker.
Several styles can share one pod. Generate with a style that has no pod opens a box of your pods and presets - + puts the style on one, or Create pod makes a new one. 📦 Presets in the PODS docker keeps a pod's template, GPU, disks and styles as a file: ▶ starts it (a terminated pod is made again with every style's files), ⏹ stops it, 🗑 terminates the pod and ✕ deletes the preset - with the hours used and what it cost.
Basic Settings' Override style works on pods too: its steps, seed, CFG and size are used, and its checkpoint is sent to the pod before the render if the pod does not have it.
Stop or terminate? A running pod is billed every hour, even when idle. ⏹ Stop ends the GPU charge but keeps /workspace - its storage is still billed. Terminate deletes the pod and every file on it and ends all charges; bitGen asks first. 🅱 marks the pods bitGen created or started.
🖥 My pods lists the pods bitGen made or started; the other pods on your RunPod account are one line under them, with what the running ones cost and ⏹ Stop - so nothing runs unseen. The Remote generation docker (Basic Settings ▸ GPU ▸ ☁ Remote generation docker) shows where the work runs.
Planned - not in the studio yet: Civitai and Hugging Face downloads sent straight to a pod; your own checkpoints and LoRAs found by their link before they are uploaded; a list to tick which of them go onto a new pod; a choice to keep, stop or terminate the pod when you switch style; bitGen terminating the pods it started when it closes; and your RunPod balance with a cost report you can save as a PDF.
OpenRouter: the LLM
From version 2.753.0 the studio's LLM - the screenplay writer, prompt help and picture descriptions - can run on OpenRouter: one API key, hundreds of models, some of them free. Nothing to install and no remote GPU needed. Pictures, videos and songs still render on your computer, and its graphics card no longer has to share room with the LLM.
- Get a keyMake a free account at openrouter.ai and create a key (Keys).
- Settings ▸ Remote generation ▸ 🧠 LLMPick OpenRouter and paste the key - it is kept encrypted with your other keys.
- Pick the modelsThe LLM, and a vision LLM for pictures (one that takes pictures, marked 👁). Free models only is ticked at first; untick it to see the paid ones with their price per million tokens. ✓ Test checks the key and the model.
- Basic Settings ▸ GPUPick Both and tick LLM and Vision LLM.
Free models allow about 20 requests a minute and 50 a day - 1000 a day once your OpenRouter account has bought $10 of credits. OpenRouter can change these limits. Your prompts and pictures go to OpenRouter and the model's provider, so keep private material on your own computer.
Pictures and video on OpenRouter, paid per picture or per second of video, came in 2.757.0 - see Online services.
Online services: Settings ▸ API
From version 2.757.0 bitGen Studio can render without a good graphics card at all: an online service does the work and charges your account for each picture, video or song. No ComfyUI and no models to download - add the service's API key and its styles appear in the style lists.
| Service | What it makes in bitGen |
|---|---|
| OpenRouter | Pictures, video, the LLM |
| fal.ai | Pictures, video, songs |
| Replicate | Pictures, video, songs |
| OpenAI | Pictures, the LLM |
| Google Gemini | Pictures, the LLM |
| Claude (Anthropic) | The LLM only |
| RunningHub | ComfyUI workflows in the cloud: pictures, video, songs |
- Get a keyMake an account with the service and create an API key. Most services need credit on the account before they render.
- Settings ▸ APIPaste the key - bitGen recognises the service from it - give it a name, leave Use ticked and press + Add. ✓ Test checks the key.
- Pick an API_ styleUse makes API_<name> pictures, video and songs styles (whatever the service makes). Pick one in the prompt docker, the video styles or the Music docker and render as usual.
- Change the modelEach style starts with a model the service offers; type another in Models and key, or in the style's Runs on ▸ API. For RunningHub the model is the workflow ID - bitGen fills in your prompt, size, seed and pictures.
Use for the LLM sends the screenplay writer, prompt help and picture descriptions to OpenRouter, OpenAI, Gemini or Claude. Keys are kept encrypted on your computer and only sent to their service. Your prompts and pictures go to the service, so keep private material on your own computer. OpenAI ended its video API in September 2026; for video use OpenRouter, fal.ai, Replicate or RunningHub.
Ready-made remote styles
From version 2.749.0 the studio comes with remote styles for RunPod's own worker images. Each image holds one set of models, and the styles use only those files and ComfyUI's own nodes - nothing to upload.
| Style | RunPod image |
|---|---|
| R_SDXL, R_SDXL img2img | runpod/worker-comfyui:5.10.0-sdxl (tested) |
| R_FLUX.1 schnell | …-flux1-schnell |
| R_FLUX.1 dev | …-flux1-dev |
| R_FLUX.1 dev fp8 | …-flux1-dev-fp8 |
| R_SD3 Medium | …-sd3 |
| R_Z-Image Turbo | …-z-image-turbo |
The studio asks RunPod which image your endpoint runs (Settings ▸ Remote generation ▸ RunPod Serverless ▸ Your endpoint's image) and lists only the styles for it - with the sdxl image, R_SDXL and R_SDXL img2img. They show when Basic Settings ▸ GPU is on Remote or Both.
Endpoints for all your styles
Planned - not in the studio yet. RunPod's API lets an app create endpoints with your API key, so the studio can set them up for you. Here is how it is meant to work.
One endpoint per style would not work well for most styles. A serverless worker can only use models that are already on it - built into its image or on a network volume - and RunPod's ready-made images hold just six model sets. Every other style would need its own image built with its models, many gigabytes each.
- One button for RunPod's own imagesThe studio makes an endpoint for each of RunPod's six ready-made images and connects each ready-made remote style to its own. With no active workers an endpoint costs nothing while it waits, and nothing needs uploading.
- One shared endpoint for every styleOne network volume holds every model your remote styles use. A short-lived pod downloads them from Hugging Face (the studio knows each style's files) and stops; one worker image carries your node packs; the endpoint gets the volume, and every remote style points at it.
Costs and limits
- RunPod charges for the GPU while a render runs, and a monthly fee per GB for a network volume - see runpod.io/pricing.
- A stopped pod costs nothing for its GPU, but its storage is billed until you terminate it.
- A network volume ties the endpoint to one data centre, so fewer GPUs are on offer to it.
- Video is heavy: MiniMax H3's models take about 70 GB of storage, and it wants a card with 20 GB or more (16 GB in 16 GB mode).
- Styles that need node packs (GGUF, LTX, Viggle) need your own worker image. RunPod can build it from a GitHub repository, once you connect GitHub to your RunPod account - that step is yours.
- Serverless: a render and its input files must be under 10 MB together, and the first render after a quiet spell waits for a worker to start.
- Your own computer's models are not used - every model must be on RunPod first.
Questions? Ask in the Discussions.