Run your own image server

MyLLM makes pictures and video on a ComfyUI server that you run. This page shows what to install so it works.

For MyLLM 6.0.1 on iPhone, iPad and Mac · updated 20 October 2026

Short version: install ComfyUI, put the FLUX 3 Image Klein files (or the FLUX.2 Klein 4B files) in its models folders, then enter the server's address in Settings → Image & Video Generation and tap Test ComfyUI. It tells you which model it found, or which files are missing.

1. What you need

A computer that stays on while you make pictures: a PC with an NVIDIA graphics card, or an Apple Silicon Mac.
About 8 GB of graphics memory for FLUX 3's int8 build or FLUX.2 Klein 4B. More memory is faster and allows the full-size builds.
15–25 GB of free disk space for the model files.

MyLLM checks which model your server has each time and shows the matching controls. You can switch models on the server later without changing anything in the app.

2. Install ComfyUI

ComfyUI is free and open source. Install it from comfy.org/download (desktop app for Windows and Mac) or from its GitHub page, and keep it up to date: the FLUX 3 nodes need a recent version.

The model files below go in the folders inside ComfyUI's models folder: diffusion_models, text_encoders and vae. Restart ComfyUI after adding files.

3. Model option A: FLUX 3 Image Klein

The newer model, with instruction photo edits ("make it a rainy night") and an optional prompt upsampler that powers Enhance prompt. The files are on Black Forest Labs' flux-3-image-klein-comfy page. You need a Hugging Face account and must accept their licence there first. It is the FLUX Kommunity licence, non-commercial unless BFL says otherwise, so read it.

Folder	File	Notes
`diffusion_models`	`flux3-image-klein-int8-convrot.safetensors`	Runs on 8 GB cards. Use `-nvfp4` on an RTX 50-series (Blackwell) card, or `-bf16` with lots of memory. Not the `-base-` files.
`text_encoders`	`qwen3vl_4b_int8_convrot.safetensors`	Required. The `bf16` build also works.
`vae`	`flux2-vae.safetensors`	Required (it is shared with FLUX.2).
`text_encoders`	`qwen3.5_9b_flux3_prompt_upsampler_w4a8.safetensors`	Optional. Turns on Enhance prompt and sharper edits. Without it, prompts go to the model as you wrote them.

Comfy's announcement, with their own workflow templates: FLUX 3 Image Klein in ComfyUI.

4. Model option B: FLUX.2 Klein 4B

The previous model. It is lighter, still good, and what MyLLM used before 6.0.1. Photo edits redraw the picture from your prompt, with a Strength slider. File names and links follow ComfyUI's FLUX.2 Klein tutorial:

Folder	File	From
`diffusion_models`	`flux-2-klein-4b-fp8.safetensors`	black-forest-labs/FLUX.2-klein-4b-fp8
`text_encoders`	`qwen_3_4b.safetensors`	Comfy-Org/flux2-klein-4B (`split_files/text_encoders`)
`vae`	`flux2-vae.safetensors`	Comfy-Org/flux2-dev (`split_files/vae`)

Check each model card for its licence before any commercial use. If both models are installed, MyLLM uses FLUX 3.

5. Video (optional)

Video uses Wan 2.2 TI2V-5B: wan2.2_ti2v_5B_fp16.safetensors in diffusion_models, umt5_xxl_fp8_e4m3fn_scaled.safetensors in text_encoders and wan2.2_vae.safetensors in vae, plus the Video Helper Suite custom nodes. It is much heavier than pictures: 16 GB or more makes it practical. See the catalogue page for the full hardware table.

6. Reaching it from your phone

At home, same Wi-Fi: start ComfyUI with --listen so other devices can reach it, then use http://<computer's address>:8188. Plain HTTP is fine on your own network only.
Away from home: don't expose ComfyUI to the internet as it is, because it has no password. Use a private network such as Tailscale (its tailscale serve gives you an HTTPS address only your devices can reach), or put it behind a reverse proxy that requires a bearer token. MyLLM sends the token from Settings as Authorization: Bearer ….

7. Connect MyLLM

Open Settings → Image & Video Generation.
Enter the server's address and, if your proxy needs one, the token. The token is kept in the iOS Keychain.
Tap Test ComfyUI. You should see "Connected to ComfyUI. Images: FLUX 3 Image Klein" (or FLUX.2 Klein). If files are missing, it names them.
Make pictures from the Create screen, ask for one in chat ("draw me a lighthouse at dusk"), or let MyLLMos apps use it.

8. Your own workflow

Prefer your own pipeline? Build it in ComfyUI, then use Save → Export (API Format) and paste the JSON into Settings → Image & Video Generation → Custom workflow JSON…. MyLLM always uses a pasted workflow as it is. It fills in the prompt, seed and size by node number. The node numbers for each model are listed under that field in the app.

9. Troubleshooting

"Your ComfyUI server doesn't have an image model MyLLM can use": A file from the tables above is missing or in the wrong folder. The message lists which ones. Restart ComfyUI after adding them, then tap Test ComfyUI again.
"Unauthorized — check your ComfyUI bearer token": Your proxy rejected the token. Copy it again into Settings.
"ComfyUI rejected the workflow": Usually an older ComfyUI that doesn't have the FLUX 3 nodes yet. Update ComfyUI. With a custom workflow, check that its node numbers match.
No Enhance prompt switch: The optional prompt upsampler file isn't installed. Everything else works without it.
The first picture takes much longer: ComfyUI loads the models into memory on the first request. Later pictures skip that.