SRPO at 1.0 dominates. Lower SRPO to 0.6-0.7 and raise the style LoRA.

Trends

.ai

de bästa AI-modellerna med öppen källkod

Flux SRPO LoRA: better skin and portraits for FLUX.1 Dev

Q: Faces look over-detailed

Guidance too high with the LoRA at 1.0. Guidance 2.5 or LoRA at 0.8.

The SRPO LoRAs apply Tencent's preference-optimized Flux weights as a LoRA, improving skin detail, lighting and portrait realism. Learn weights, settings and prompt ideas for photographic people.

Ämne: Modeller

Av Captain

Publicerad 2026-09-30

Den här sidan har inte översatts ännu, så den visas på engelska.

Översikt

SRPO (Semantic Relative Preference Optimization) is a fine-tuning method that aligned

FLUX.1 Dev

toward human preferences, dramatically improving realism in portraits and skin. These

LoRA

extractions by Alissonerdx bring that improvement to any Flux Dev workflow instead of requiring the full

checkpoint

Use it as a realism base layer: it produces better portraits and skin than Dev or Krea alone and stacks with character LoRAs.

Referens

Namn	Typ	Vad det är
LoRA weight	0.7-1.0	1.0 for full effect; 0.7 when stacking a character LoRA.
Trigger	none	None needed.
Guidance	2.5-3.5	Lower values keep skin natural.
Steps	25-30	Standard Dev.
Base	FLUX.1 Dev	Dev or Krea Dev.

Inget att installera - kör AI-modeller i molnet

BitVector Prism är det enklaste sättet att komma igång: välj en modell, skriv en prompt och generera i en ren webbapp, utan någon installation alls. BitVector finns också på Discord och på webben (SpyGlass).

Prova BitVector

Steg för steg

Attach SRPO at 1.0 to Flux Dev.
Write a photographic prompt with a described light source.
Set guidance 3, 28 steps.
Compare the same seed with and without the LoRA; look at pores, hair and light falloff.
Add a character or style LoRA at 0.7 if needed and lower SRPO to 0.8.

Exempel

Window-light portrait

A close portrait of a man in his fifties with a salt-and-pepper beard by a north-facing window, natural skin texture, 85mm, f/2.

SRPO 1.0, guidance 3

Group scene

Three friends laughing around a kitchen table at dusk, warm pendant light, candid, 35mm.

SRPO 0.9

Creative idea: period portrait

A 1940s-style studio portrait of a jazz singer, single hard key light, film grain, black and white.

SRPO 1.0

Tips

Natural light prompts (window light, overcast) show the biggest gains.
It also improves lighting in non-portrait scenes; try it on interiors.
Keep guidance low; SRPO's realism and high guidance fight each other.
Pairs well with the ULP amateur-photo LoRA for candid looks.

Felsökning

Faces look over-detailed

Varför det händer

Guidance too high with the LoRA at 1.0.

Så löser du det

Guidance 2.5 or LoRA at 0.8.

Style LoRA weakened

Varför det händer

SRPO at 1.0 dominates.

Så löser du det

Lower SRPO to 0.6-0.7 and raise the style LoRA.

Inget att installera - kör AI-modeller i molnet

PirateDiffusion finns bara på Telegram och är byggt för proffs: tusentals modeller, LoRA och arbetsflöden som styrs med chattkommandon, med obegränsad generering till ett fast pris.