Start with the matching official workflow
ComfyUI documents full and FP8 FLUX.1 examples. Use the template for your chosen variant; different packages need different files. Check the listed download and license requirements before installation.
Verify files before adding adapters
For the documented full workflow, check the diffusion model, text encoders and VAE against the official guide. Confirm that each loader selects the intended file. Do not assume a community quantization uses the same loader as the official package.
Full FLUX.1 dev: which file goes where?
For the full dev workflow, ComfyUI's guide separates the diffusion model, two text encoders and VAE. A single-file FP8 checkpoint is a different packaging path: use its matching template rather than combining instructions. The table below is for the documented full workflow only. Refresh the model list or restart after moving files.
| Component | File in the full dev example | Folder |
|---|---|---|
| Diffusion model | flux1-dev.safetensors | ComfyUI/models/diffusion_models/ |
| Text encoder | clip_l.safetensors | ComfyUI/models/text_encoders/ |
| Text encoder | t5xxl_fp16.safetensors | ComfyUI/models/text_encoders/ |
| VAE | ae.safetensors | ComfyUI/models/vae/ |
Dev, schnell and model rights
Black Forest Labs identifies dev as an open-weight model governed by its FLUX.1 dev non-commercial license. Its card separately describes permitted use of generated outputs; do not confuse output rights with unrestricted commercial operation of the model. Schnell's card lists Apache-2.0 and one-to-four-step generation. Those publisher descriptions do not prove which variant is best for your project. Check the exact artifact and license before choosing a workflow.
Measure the complete pipeline
Record model files, precision, resolution, batch size, ComfyUI version and offload settings. Save a baseline output and peak memory measurement. If a run fails, change one setting and preserve the error log. No hardware benchmark is reported here.
Add one compatible LoRA after the baseline
Save the working graph, confirm adapter compatibility, then compare the same input with and without the adapter. Follow the adapter's instructions and license. Training a new LoRA is a separate workflow; these inference checks do not establish training-memory requirements.
Continue with safe installation
Use the model download safety checklist before installing unfamiliar files, and the ComfyUI first-image guide when troubleshooting a basic graph.
Troubleshoot the graph before installing more nodes
Use a small, saved baseline and change one variable per retry. The following is a diagnostic checklist, not a record of tests performed on our hardware.
| Symptom | Next check |
|---|---|
| A model is missing from a loader | File folder, extension and loader selection; refresh the model list |
| Full and FP8 instructions conflict | Return to the official template for the exact package |
| Out of memory | Reduce the workload and inspect allocation; do not infer fit from file size |
| Baseline works but LoRA fails | Remove the adapter; verify family, loader and adapter instructions |
| Output quality changes after optimization | Compare the identical prompt and settings, retaining both images |
If the workflow runs out of memory
Use the ComfyUI memory-error guide to separate loading, sampling and VAE failures. Record one change per retry in the downloadable log rather than combining unrelated launch flags.
Frequently asked questions
Is there one minimum VRAM figure for every FLUX workflow?
No. Model precision, text encoders, output size, batch size, runtime and offloading affect memory use. A checkpoint's download size is not the complete pipeline's memory requirement.
When should I add a LoRA to a FLUX workflow?
First save a working base workflow and its output. Check that the adapter matches the model and loader, then add one LoRA and compare using the same input and settings.
Where do FLUX.1 dev model files go in ComfyUI?
In the documented full workflow: the diffusion model goes in models/diffusion_models, CLIP-L and T5 text encoders in models/text_encoders, and ae.safetensors in models/vae. Other packages require their matching template.
Is FLUX.1 dev the same as schnell?
No. They have separate checkpoints, sampling guidance and licenses. Check the model cards and use the workflow for the chosen variant.
Sources & further reading
Primary sources checked Oct 26, 2026. Vendor statements are attributed; editorial advice is our own.
- 1
- 2FLUX.1 dev model card and license ↗Black Forest Labs
- 3FLUX.1 schnell model card ↗Black Forest Labs
Help us keep this useful. Report a factual correction →
