realrebelai/Kreamagine_v1.0

Kreamagine

A finetune of Krea-2-Turbo built around a Grok Imagine derived style adaptation, released in INT8 ConvRot and the full GGUF ladder so it runs on 8GB cards.

Krea-2-Turbo is a 12.9B single-stream DiT distilled for 8-step generation. Kreamagine keeps that step count and sampler behaviour intact — it is a style and rendering finetune, not a re-distillation.

What it changes

The adaptation was trained on a dataset assembled from Grok Imagine v0.9 and v1.0 generations, then merged into the Turbo base.

What you actually get out of it:

• Lighting. The most pronounced difference. Much more aggressive light shaping — harder key lights, stronger falloff, deeper contrast ratios, more willingness to let regions go fully black or blow out. Base Krea-2-Turbo tends toward even, safe illumination; this does not.
• Detail density. Noticeably more micro-detail in surfaces, fabric, skin and foliage. Fine structure holds together at 1024–1536 rather than smoothing out. Detail does see hallucinations past 1536 resolution so keep it under there for best results.
• Aesthetics, prompt-dependent. There is a stylistic pull, but it is not baked in uniformly. Neutral or descriptive prompts stay close to base behaviour. Prompts that lean into mood, cinematography or lighting language pull the aesthetic much harder. If you want the base look, prompt plainly; if you want the shift, prompt for it.

No trigger word. The adaptation is merged into the weights.

Files

INT8 ConvRot

Kreamagine-INT8-ConvRot.safetensors

Row-wise INT8 with ConvRot (group size 256). Loads with the standard Load Diffusion Model / UNETLoader node — no custom nodes needed. Rotation metadata is embedded per-layer and read automatically.

Requires ComfyUI v0.27.0 or newer (native INT8 support) and a current comfy-kitchen. On Ampere cards (RTX 30 series) this is generally the one you want: those GPUs have INT8 tensor cores and no FP8 hardware, so INT8 beats FP8 outright, and ConvRot closes most of the remaining gap to bf16.

The input/output projections, timestep MLP, text-fusion stack and modulation path are kept at full precision. Only the 28 transformer blocks are quantized.

GGUF

Tier	Notes
Kreamagine-Q8_0.gguf	Highest fidelity of the ladder. Closest to bf16.
Kreamagine-Q6_K.gguf	Very close to Q8 at a useful saving.
Kreamagine-Q5_K_M.gguf	Good balance.
Kreamagine-Q4_K_M.gguf	The 8GB sweet spot.
Kreamagine-Q3_K_M.gguf	For 6GB cards or heavy concurrent VRAM use.
Kreamagine-Q2_K.gguf	Smallest. Expect visible degradation.

Quantized from a single pass off the bf16 merge — no double quantization, and no FP8 or INT8 intermediate.

Workflow

A ready-to-run ComfyUI workflow is included in this repository. Drop the JSON into ComfyUI and point the loaders at your files.

Setup

ComfyUI/
├── models/
│   ├── diffusion_models/
│   │   └── Kreamagine-INT8-ConvRot.safetensors   (or a .gguf tier)
│   ├── text_encoders/
│   │   └── qwen3vl_4b_fp8_scaled.safetensors
│   └── vae/
│       └── qwen_image_vae.safetensors

Text encoder and VAE are unchanged from stock Krea-2 — grab them from Comfy-Org/Krea-2 if you don't already have them.

Sampler settings

Setting	Value
Steps	8
CFG	1.0
Sampler	euler
Scheduler	simple
Shift	1.15
Resolution	1024–1536

CFG 1.0, not 4.0. Turbo is distilled and has no classifier-free guidance — raising CFG will not improve adherence, it will just wreck the output. Negative prompts are ignored at CFG 1.0.

Prompt in natural language with detailed description. Because the lighting response is strong, lighting and camera terms carry more weight here than on base Turbo — that is the main lever worth learning.

License

Governed by the Krea 2 Community License Agreement, inherited from krea/Krea-2-Turbo. That license flows through to finetunes and quantizations, including this one, and it restricts commercial use. Read it before you deploy anything built on these weights.

Report commercial misuse of Krea 2 derivatives to opensource@krea.ai.

Credits

Base model by Krea AI. Reference-format weights repackaged by Comfy Org.

Finetune, quantization and ComfyUI integration by RealRebelAI — GitHub · Hugging Face · X