Design & Media: Creating Realistic Digital Imagery using Krea-2-Turbo / Z-Image Turbo

Creative Stack & Specs

• Core Toolset: Krea-2-Turbo (FP8 or BF16) + Z-Image Turbo (FP8), SGLang Diffusion
• Target Medium: Fast-turnaround digital images (~3s per gen; any resolution including 2mp target)
• Key Parameters (Krea 2): steps=13, cfg_scale=1.0, Euler sampler, Simple scheduler

Step-by-Step Production Workflow

1. Environment Setup/Installation: Install the required library for advanced conditioning using uv pip install 'sglang[diffusion]' --prerelease=allow.
2. Model Selection based on Hardware Constraints:
   • For higher quality if hardware permits: Use Krea-2-Turbo (approx. 13gb).
   • If running weaker hardware: Select Z-Image Turbo (approx. 6gb - faster and lighter weight).
3. Uncensoring via Prompt Rebalancing / Conditioning: To bypass content restrictions in Krea-2-Turbo without relying solely on LoRAs, apply a rebalancer parameter to sglang diffusion by passing post-prompting parameters such as [e.g.,] "1, 1, 1, 1, 1, 1, 1, 2.5, 5.0, 1.1, 4.0, 1.0" with a multiplier of 2 at the generation endpoint.
4. Advanced Image Editing/Stylizing (Optional): Utilize reference images or masks allowed within these models to perform image editing rather than just text-to-image prompting.

Style Consistency & Quality Controls

• Censorship Bypass Strategy: Use specifically trained weights like krea2filterbypass3 (.safetensors) which is under 1kb for improved prompt adherence으로 avoiding blockers; alternatively use heavy conditioning if using SGLang.
• Anatomical Correction (Hands/Fingers): Monitor output carefully during training and selection phases, especially when doing certain styles (like anime), though both models show improvement over SDXL in handling hand anatomy errors.
• Negative Prompt Control: Implement NegPiP node logic where negative sentences can be used via general strength and token strength settings kept at -1.
• Texture Enhancement: For skin texture realism, consider applying light custom LoRAs such as those ablevated by low step counts (~200 steps).

The bottom line: This pipeline enables high-speed (~3s per image) photorealistic asset creation with flexible censorship control through sglang rebalancing or specialized lightweight filters.

! DYOR (Do Your Own Research)