Easy Krea 2 variance in ComfyUI: reuse the text encoder as an LLM for prompt expansion, no LoRAs needed
prookyon · reddit · 2026-10-03
Responding to complaints that Krea 2 outputs are too predictable, the author shares a simple trick: reuse ComfyUI's text encoder model (e.g., Qwen3VL 4B) as an LLM for prompt expansion (PE) — no LoRAs or model-specific custom nodes needed.
Key points:
- Any image-generation-targeted system prompt works; the author even uses Qwen 2.1's PE system prompt. The official ComfyUI template already has PE — you can swap in any prompt you like.
- You get two knobs: PE seed for drastic changes, sampler seed for fine-grained variation.
- Minimal overhead: 18s for a 3MP image (int8, 8 steps) on a 5070Ti, with PE adding only 4s; the text encoder loads anyway, so no extra VRAM pressure.
- No need for abliterated/uncensored encoder versions — the normal Qwen3VL 4B is surprisingly permissive as a text generator, and alternative encoders hurt image quality.
- Custom nodes in the workflow (rgthree, easy-use, etc.) are purely for convenience, not part of the PE trick.
The post includes the first 5 unfiltered random results from a tiny prompt, plus a full workflow link.
More from Multimodal
- Reddit user generates a convincing Back to the Future IV teaser with AI video tools — chodtoo · 2026-10-03
- Solo dev builds Chimera Arena, an AI card battler generating turn-by-turn battle videos from open models — Ill-Ant-9489 · 2026-10-03
- ComfyUI-Omnichar ships .char format for consistent characters across Flux Klein and MiniMax H3 — ashishsanu · 2026-10-03
- Redditor shares weeks-long deep dive into how WAI Illustrious actually works — Limp_Okra_5496 · 2026-10-03
- Temporal Ghost Imaging prompt: time-shifted color channels create phantom AI images — LudovicCreator · 2026-10-03
- Cutting AI Video Costs with 70s Cel Animation Techniques, via Opus — Aargau · 2026-10-03