ComfyUI beginner asks how to separate pose, identity and background references
butterfly--1313 · reddit · 2026-09-29
A ComfyUI Cloud beginner details two use cases on Reddit: recreating a Pinterest reference photo with herself as the subject (preserving pose, composition, lighting, and realistic skin texture), and a harder variant adding separate environment and clothing references.
Experiments with Qwen Image Edit 2511 suffer identity drift, proportion changes, and over-smoothed "AI glamour" results. She suspects a workflow/conditioning problem rather than a prompting one, and asks: best current model combo (nano banana pro? Flux Kontext?), whether FaceID/InstantID/PuLID or identity LoRAs are still the right approach, how to transfer pose without inheriting the original person's identity, and whether references can be explicitly role-assigned. Includes a basic workflow screenshot.
More from Multimodal
- Remaking a 1987 Amiga game intro locally with Flux 2 Klein, Qwen Image Edit, MiniMax H3 and YuE2 — astrofiles · 2026-09-29
- Claude generates a striking 2-minute video of Google's history — FinanceYF5 · 2026-09-29
- A reusable Celestial Aurora Borealis prompt template with swap-in colors — LudovicCreator · 2026-09-29
- Doctor generates a full Switzerland travel vlog in 20 minutes with ChatGPT — DrDatta_AIIMS · 2026-09-29
- Topaz Labs ships faster video upscaling on web, SDR-to-HDR in Premiere, and Android debut — azed_ai · 2026-09-29
- Paste a URL, Get a Manus-Style Launch Video: Skill Built on Claude Opus 5.5 — yihui_indie · 2026-09-29