Report: working on Apple Silicon / MPS — M4 Pro, 48GB (incl. dual masked references + upstream LLM prompt input)

[talib79g-cmyk]

Not a bug — a validation report, since I couldn't find any Apple Silicon results for this node and thought it might help other Mac users deciding whether to download the bf16 encoder.

Setup

Mac Mini M4 Pro, 48GB unified memory
macOS 27 developer beta
ComfyUI v0.26.2, Python 3.11.14 venv, PyTorch Your ComfyUI-SCAIL venv has PyTorch 2.14.0.dev20260624 — a nightly build from June 24, 2026.
Diffusion model: krea2_turbo bf16 (fp8 diffusion avoided on MPS generally)
Text encoder: qwen3vl_4b_bf16.safetensors (8,875,719,384 bytes) via CLIPLoader type krea2

Confirmed working on MPS

Text-only path (baseline, matched stock CLIP Text Encode output quality)
Single reference image through the vision path
Reference masking (MaskEditor face mask → mask input; scene contamination from the reference photo's background disappeared as documented)
Two reference images (male + female), each with its own face mask — Picture 1/Picture 2 labelling appears in print_prompt as expected, and prompt gender terms mapped correctly to the right reference without explicit Picture numbering
Prompt text supplied via a connected upstream LLM node (stavsap comfyui-ollama) rather than a widget
print_prompt, default system prompt, vision_position=before prompt

Your fp8 error path is accurate — the fp8_scaled encoder is indeed text-only; everything above required the bf16 file.
Memory behaviour (48GB)
With the ~24GB bf16 diffusion model resident, the vision encode step shows transient yellow memory pressure in Activity Monitor (brief yellow also occasionally during sampling), green otherwise, with both one and two references. No slowdown, no errors across many runs.

One user-error trap worth noting for others: since system_prompt is a connectable input sitting next to the prompt input, I initially wired my LLM's output into system_prompt by mistake — print_prompt then shows your text (in the system slot) so it looks connected, but the user section is empty and generations ignore the prompt. print_prompt made the diagnosis easy once I knew to check which section the text was in.
Not yet tested on my side: the Krea2 System Prompt instruct mode and vision_position=after prompt.
Thanks for building this — the masking design in particular works exactly as the README describes.

[talib79g-cmyk]

Follow-up after further testing on the same setup (M4 Pro 48GB, MPS, bf16 encoder), specifically on two-reference mode — a finding that may be useful for the README or for anyone hitting inconsistent results.
TL;DR: with two masked references, integration vs. failure appears to be seed-dependent, not prompt-dependent.
The failure mode: on some generations, instead of placing the two referenced people into the prompted scene, the model paints the reference crops themselves as giant close-up panels/faces composited around a fragment of the actual scene — a collage of the references rather than a scene containing them. Likeness is preserved inside the panels; prompt adherence collapses. Removing words like "photo/portrait/reference" from the prompt did not help.
Controlled tests that isolated the variable (all on the same graph, two masked face references, default descriptor system prompt, vision_position=before prompt):

Re-running a previously successful generation from its saved PNG (same seed, same prompt) reproduced the good, integrated result almost exactly — so the pipeline itself is deterministic and intact.
Keeping that known-good seed but swapping in the full text of a prompt that had just produced collages → good integrated image. The failing prompt is exonerated on the good seed.
The same prompts on other seeds → collage.

Conclusion: some seeds resolve the side-by-side text+vision conditioning into scene integration, others into depiction of the references, largely independent of prompt wording. Practical workaround that's serving me well: treat clean-integrating seeds as reusable "golden seeds" and vary only the prompt on them; hunt for additional ones via randomized batches.
Also tested: the Krea2 System Prompt instruct node in the two-reference case — negative result. Both the shipped instruction and an edited variant (explicitly framing the pictures as identity references for a woman/man and forbidding depiction of the references themselves) produced worse collages than the default descriptor, including an extra unprompted person — consistent with your out-of-distribution warning. For two references, the unconnected default has been the better behaviour in my testing.
Single-reference mode remains solid here.