KREA 2 Turbo: Photorealism Settings & Prompt Framing Guide

TL;DR

• Use er_sde sampler with simple scheduler at 4–9 steps, CFG 1.0 for clean KREA 2 Turbo output; the widely cited res_2s/beta sampler does not exist in ComfyUI.
• Control framing by word order: describe the room first for wide shots, the subject first for portraits — the model allocates visual "budget" to whatever gets the most words earliest.
• KREA 2 Turbo's FP8 quantization (51.5 % smaller) preserves photorealism while running on 16 GB VRAM; it outperforms peers on landmark recognition but trails on painterly styles.

Why This Model Matters

KREA 2 Turbo arrived in June 2026 as a distilled 8-step variant of the KREA 2 Raw base model. Its claim to attention is simple: a 20-billion-parameter diffusion transformer compressed to 12 GB via selective FP8 quantization while keeping photorealistic quality intact. The quantization preserves 166 sensitive tensors — including the final-layer modulation vectors — in full precision, avoiding the promotion bugs that usually degrade FP8 output.

For practitioners, this means a model that fits on a 16 GB GPU and still renders the Oslo Opera House correctly, reproduces hands holding keys without mangled fingers, and — unusually — reproduces exact meme templates pixel-for-pixel rather than improvising look-alikes.

Verified Technical Stack

Note: The CivitAI article that popularized res_2s/beta appears to reference a custom or hypothetical sampler. ComfyUI's KSAMPLER_NAMES registry contains er_sde but no res_2s variant. Until an official implementation surfaces, treat res_2s/beta as unverified.

Photorealism Settings That Actually Work

Recommended Baseline

Sampler: er_sde
Scheduler: simple
Steps: 8 (Turbo default) or 4–9 for quality/speed trade-off
CFG: 1.0
AG: 0.0
Resolution: 1024×1024 or 2048×2048 (native)
Text Encoder: Qwen3-VL-4B FP8 (Krea2TEModel)
VAE: Qwen-Image VAE

Why These Settings Hold Up

• CFG 1.0 / AG 0.0 matches the distillation target; higher values introduce artifacts without fidelity gains.
• er_sde + simple is the only sampler pair verified in ComfyUI's source. The simple scheduler produces a linear-in-log-space sigma schedule that aligns with the Turbo distillation.
• 4–9 steps covers the practical range: 4 steps for preview, 8 for default, 9 when you need marginally cleaner skin texture.

Latent Upscale Workflow

The community workflow (RLY v003) adds a basic latent upscale after the initial 8-step generation. This is a standard 1.5×–2× latent upscale with a second 4-step er_sde pass — not a distinct model or sampler.

Prompt Framing: Word Budget = Zoom

The most reproducible framing technique for KREA 2 (and likely other DiT models) treats the prompt as a spatial budget. The frame belongs to whatever you describe first and most.

Practical Rules

1. Order decides priority. Room-first = wide. Subject-first = tight.
2. Word count decides dominance. The element with more descriptors claims more pixels.
3. Anchors beat crop labels. "Feet touching the carpet" pulls feet into frame; "full body shot" alone often fails.
4. Contact points set scale. "Hand on desk," "sitting back in chair" — these spatial relationships lock the subject to the environment.

The technique was documented for KREA 2 by Loraholic (CivitAI 32410) and appears transferable to other diffusion transformers because it exploits the universal cross-attention scheduling of text conditioning.

How KREA 2 Turbo Compares to Other 2026 Models

*Community-reported VRAM estimates; not vendor-published. CivitAI 31823 provides the KREA vs. Nano Banana vs. ERNIE comparison images.

Where KREA 2 Turbo Wins

• Landmark accuracy: Oslo Opera House, specific interior architectures.
• Hand anatomy: Keys, glasses, natural poses — fewer fused fingers than peers.
• Meme reproduction: Pixel-perfect template recall (Woman Yelling at Cat, Three Spidermen).
• Local deployment: Full stack runs offline on 16 GB VRAM.

Where It Falls Short

• Painterly / illustrative styles: Oil, watercolor, ASCII — clearly behind Nano Banana 2 and ERNIE.
• Abstract composition: Less inventive on non-representational prompts.
• Ecosystem maturity: Fewer public LoRAs, no official ControlNet adapters yet.

Common Errors and Fixes

Sources

FAQ

What sampler should I use for KREA 2 Turbo photorealism? er_sde with the simple scheduler at 4–9 steps, CFG 1.0. The res_2s/beta sampler cited in some tutorials is not present in ComfyUI.

Does the "word budget" framing trick work on other models? The mechanism — early conditioning tokens receiving stronger cross-attention — is architecture-agnostic for diffusion transformers. Community reports suggest it transfers to Flux, SD3.5, and Qwen-Image, but controlled comparisons are lacking.

Can KREA 2 Turbo run on 12 GB VRAM? The FP8 checkpoint is 12.01 GiB. With ComfyUI's model offloading and fp8 compute, 12 GB cards can generate at 1024²; 16 GB is comfortable for 2048².

Is the meme reproduction memorization or generalization? Unverified. The model reproduces exact template pixels for widely distributed memes (Woman Yelling at Cat, Three Spidermen). Whether this is training-set memorization or structural generalization has not been tested under controlled conditions.

What license applies to commercial use? The KREA 2 License Agreement governs all variants. It is a custom license — not Apache-2.0 or MIT. Review the full text at krea.ai/krea-2-licensing before commercial deployment.

How does KREA 2 Turbo handle text in images? Strong for short phrases (signage, UI mockups), weaker on paragraph-length rendering compared to GPT-4o Image or Nano Banana 2.

Internal Links (Placeholders)

• ComfyUI sampler deep-dive
• Prompt engineering for diffusion transformers
• FP8 quantization trade-offs
• 2026 model benchmark methodology
• KREA LoRA library overview

Image Placeholders