ComfyUI error: CUDA out of memory on load

Set weight_dtype to fp8_e4m3fn in Load Diffusion Model, or use the fp8 single-file checkpoint (source tip).

ComfyUI error: System RAM exhausted when loading T5

Use t5xxl_fp8_e4m3fn_scaled.safetensors; source recommends fp16 only with more than 32GB RAM.

Flux.1 Dev text-to-image

High-quality text-to-image with Black Forest Labs' Flux.1 Dev (guidance-distilled, ~20 steps).

Estimated VRAM

~23.8 GB — ⚠️ ESTIMATE (unverified real-world usage). Estimate, not a measured figure: weight-size floor = flux1-dev.safetensors (23.8 GB, sizes verified on Hugging Face/GitHub). Activations add more; ComfyUI offloading can reduce it. Source, checked 2026-10-13.

Models required

flux1-dev.safetensors — 23.803 GB ✅ verified 2026-10-13 (listing) → models/diffusion_models
t5xxl_fp16.safetensors — 9.788 GB ✅ verified 2026-10-13 (listing) · sha256 6e480b09fae049a72d2a8c5fbccb8d3e92febeb233bbe9dfe7256958a9167635 → models/text_encoders
clip_l.safetensors — 0.246 GB ✅ verified 2026-10-13 (listing) · sha256 660c6f5b1abae9dc498ac2d21e1347d2abdb0cf6c0c0c8576cd796491d9a6cdd → models/text_encoders
ae.safetensors — 0.335 GB ✅ verified 2026-10-13 (listing) → models/vae

Tested GPUs

No verified test reports yet.

Nodes

UNETLoader DualCLIPLoader VAELoader CLIPTextEncode FluxGuidance KSampler VAEDecode

Common errors & fixes

CUDA out of memory on load: Set weight_dtype to fp8_e4m3fn in Load Diffusion Model, or use the fp8 single-file checkpoint (source tip). ✅ per source
System RAM exhausted when loading T5: Use t5xxl_fp8_e4m3fn_scaled.safetensors; source recommends fp16 only with more than 32GB RAM. ✅ per source

Low-VRAM tips

⚠️ General guidance, not individually verified.

weight_dtype fp8 halves model memory (per source)
Use fp8 T5 text encoder
Use the fp8 single-file checkpoint

Source & credit

Workflow by comfyanonymous (ComfyUI_examples). Get it from the original: https://comfyanonymous.github.io/ComfyUI_examples/flux/. License: Flux.1 Dev model: FLUX.1 [dev] Non-Commercial License.