FLUX.3 [klein] on a 12 GB RTX 5070: four tries, one clean run

FLUX.3 [klein] came out in October 2026. It needs a newer ComfyUI than the container image we use, so each rented node updates ComfyUI to v0.42.1 at start-up.
What happened
- Try 1 — failed. The model files are gated on Hugging Face. The start-up downloader in our container did not send the access token, so the download stopped with HTTP 401. Fix: download the two gated files ourselves, with the token, before ComfyUI starts.
- Try 2 — failed. Downloads worked (11.1 GB model, 8.9 GB text encoder). ComfyUI then exited about one second into the first image. Cause not found yet.
- Try 3 — worked. Same settings on another node. 1024×1024, 15 steps,
er_sde, cfg 1: 11.8 s per image (server-side median of 5 runs), 13.5 s measured from the client. Peak VRAM was about 11.4 GB of 12 GB. - Try 4 — failed. The comparison gallery (10 images) failed on every image. The log showed the 11 GB model download had failed, so ComfyUI ran without it. Fix: resumable retries and a size check against the server's file size; the node now stops instead of starting without the model.

Observations (drafted with Astra, checked by the author)
- In the side-by-side image, both portraits face the viewer beside a window. The left figure wears a cap and a zipped jacket; the right has uncovered white hair and an open checked shirt. Because the faces and clothes differ, it is hard to compare matching details.
- Both portraits are lit from the left, with softer light on the far side and blurred interiors behind. The right portrait (FLUX.3) has warmer skin tones and deeper forehead creases.
- In the single FLUX.3 image, fine beard strands, stray eyebrow hairs and skin highlights are clearly visible. Several chest hairs form conspicuous pale loops, which looks like an artefact.
- No hands or readable text appear in these images, so we cannot judge those from this note.
What we learned
- On a 12 GB card FLUX.3 [klein] fits, but only just. An 8 GB card is unlikely to work (not tested).
- It is much slower than FLUX.2 [klein] 4B (about 3.3 s on the same GPU) at its default 15 steps.
- One node is a reference value only. We want at least five nodes before calling it a result.
