PiD invents bright specular specks and ages skin on portraits — qwenimage 4-step checkpoint

Checkpoint: checkpoints/PiD_v1pt5_res2kto4k_sr4x_official_qwenimage_distill_4step/model_ema_bf16.pth

Summary

Decoding a portrait latent with PiD instead of the VAE introduces two artefacts that are not
present in the VAE decode of the same latent:

1. Hard bright specks on skin — small, high-contrast, sometimes prismatic. They cluster on
   the upper eyelid, the brow ridge and the inner corner of the eye, and read as glitter or
   sweat beads rather than as skin.
2. Apparent ageing — deep naso-labial folds, under-eye hollows and crease lines that the
   VAE decode does not contain. A prompt asking for a 25-year-old subject decodes as someone
   roughly 15 years older.

Both are generated, not amplified: the VAE decode of the identical latent at the identical
seed is clean. Measured on skin, PiD raises the small-bright-blob rate by 43–61× over that
baseline.

The decoder is doing genuinely good work elsewhere in the same image — individual eyebrow
hairs, eyelashes and iris structure are resolved convincingly. The problem is specific to skin.

Measurements

Krea 2 Turbo, q8, 512×512 generated → 2048×2048 PiD output, fixed seed 116539113, 8 steps.
"hi-freq" is Laplacian variance of the PiD output relative to a Lanczos-4× upscale of the VAE
decode of the same latent; ΔRGB is the per-channel mean difference against that same baseline.

Speck rate on skin (white top-hat residual, per million skin pixels, all compared at 2048×2048):

What does not fix it

• degrade_sigma. Raising it to 0.20 roughly doubles invented facial detail
  (68× → 138×), darkens the image monotonically, and makes the specks brighter and
  harder-edged. It also pushes the result ~3× further from the VAE decode. An independent
  sweep by another contributor found the same direction and magnitude on a different subject
  (54× → 99× at the same sigma values).
• Sampler step count. 10 steps measured slightly worse than 8 (7,850 vs 6,873), so this
  is not the base sampler under-converging.
• Lighting. Prompting for flat, diffuse, matte, no-catchlight lighting reduces the speck
  rate by ~29% (7,850 → 5,550) but leaves it at 43× the VAE baseline. Specular highlights in
  the source appear to contribute but are not the driver. Caveat: changing the prompt shifts
  the composition, so this comparison is not measuring identical anatomy and the 29% is soft.

Is it the checkpoint or the base model?

Across the three released v1pt5_res2kto4k checkpoints, on one fixed prompt and seed, the
qwenimage checkpoint is a clear outlier:

This does not isolate the checkpoint from the base model. Krea 2 is the only model tested
that routes to the qwenimage checkpoint, so "the qwenimage checkpoint over-textures skin" and
"Krea 2's latents provoke PiD" are not separated by this data. The discriminating test is
Qwen-Image itself — same checkpoint, different generator — which has not been run here.

Worth noting the non-portrait column: every checkpoint drops substantially on a non-skin
subject (a puffin on a cliff), and the qwenimage result there is clean and artefact-free.
Whatever this is, it is specific to skin rather than to the checkpoint's overall behaviour.

Images

Reproduction

prompt  a 25yo Asian women with brown eyes, sitting in a cafe with a cup of cappuccino.
        she wears a cobalt blue tank top, and the cappuccino has a flower design in the foam.
seed    116539113
size    512x512  ->  2048x2048
steps   8
model   Krea 2 Turbo, 8-bit

Decode the same latent twice — once through the VAE, once through PiD — and compare the eye
and brow region at 1:1.

Environment / caveat

These renders come from an MLX port of PiD for Apple Silicon
(filipstrand/mflux#490), not NVIDIA's
reference PyTorch implementation, so a port-side defect cannot be ruled out from this evidence
alone.

Two things argue against that being the whole story: the port reproduces the reference
behavior cleanly on the other two checkpoints and on non-skin subjects, and the same
over-texturing of skin has been reported by ComfyUI users — degrade_sigma ≈ 0.2 is the
commonly suggested remedy there, which the measurements above show makes it worse rather than
better.