AI Model Parameter & CLI Flag Calibration Hub
Master Mathematical Weights, Reference Anchors & Aesthetic Levers
Model parameters are the secret to moving beyond unpredictable AI roulette. Calibrate exact syntax, reference weights, guidance scales, and stylistic flags across Midjourney v6.1, FLUX.1 Dev, and Stable Diffusion XL with dedicated benchmarks and common mistake diagnoses.
Style Reference (--sref)
Transfers the complete aesthetic palette, color grading, brushstrokes, and visual vibe from reference images onto a new prompt subject without copying the reference subject.
Character Reference (--cref)
Maintains facial likeness, hair color, and character identity across multiple scenes, outfits, and camera angles.
Stylize Control (--stylize)
Controls how strongly Midjourney's default house aesthetic, artistic flair, and color harmony are applied versus literal prompt adherence.
Chaos & Initial Variation (--chaos)
Controls how varied and divergent the 4 initial image grid outputs are from one another.
Weird Experimental Mode (--weird)
Introduces quirky, surreal, and offbeat aesthetic properties into your generations.
Style Raw (--style raw)
Dampens Midjourney's default house aesthetic to render authentic documentary photographic realism, natural skin micropores, and uncandid textures.
FLUX Guidance Scale
Calibrates flow-matching transformer guidance strength in FLUX.1 models. Balances prompt compliance against natural contrast and skin tone fidelity.
SDXL Classifier-Free Guidance (CFG)
Controls how aggressively the U-Net latent diffusion denoising steps follow the positive text prompt versus unconditioned generation.
Parameter Quick Calibration Cheat Sheet
| Parameter Flag | Model | Default | Recommended Sweet Spot | Primary Effect |
|---|---|---|---|---|
| --sref | Midjourney v6.1 | None | --sref <URL> | Transfers the complete aesthetic palette, color grading, brushstrokes, and visual vibe from reference images onto a new prompt subject without copying the reference subject. |
| --cref | Midjourney v6.1 | None (Character Weight defaults to 100) | --cref <URL> --cw 100 | Maintains facial likeness, hair color, and character identity across multiple scenes, outfits, and camera angles. |
| --stylize | Midjourney v6.1 | 100 | --stylize 50 | Controls how strongly Midjourney's default house aesthetic, artistic flair, and color harmony are applied versus literal prompt adherence. |
| --chaos | Midjourney v6.1 | 0 | --chaos 0 | Controls how varied and divergent the 4 initial image grid outputs are from one another. |
| --weird | Midjourney v6.1 | 0 | --weird 250 | Introduces quirky, surreal, and offbeat aesthetic properties into your generations. |
| --style raw | Midjourney v6.1 | Standard Model Bias | --style raw | Dampens Midjourney's default house aesthetic to render authentic documentary photographic realism, natural skin micropores, and uncandid textures. |
| guidance_scale | FLUX.1 Dev | 3.5 | guidance_scale: 2.5 | Calibrates flow-matching transformer guidance strength in FLUX.1 models. Balances prompt compliance against natural contrast and skin tone fidelity. |
| CFG Scale | Stable Diffusion XL | 7.0 | CFG: 5.0 - 6.0 | Controls how aggressively the U-Net latent diffusion denoising steps follow the positive text prompt versus unconditioned generation. |
Frequently Asked Questions
What are AI model parameters and CLI flags?
Parameters (such as --stylize, --sref, or guidance_scale) are mathematical hyper-parameters passed to generative diffusion models that control aesthetic variance, prompt compliance, image aspect ratio, or reference image conditioning without changing the core text prompt.
Which model parameters are most important for photographic realism?
For Midjourney v6.1, using --style raw combined with moderate stylization (--stylize 50 to 150) produces the most natural, unretouched results. For FLUX.1 Dev, keeping guidance_scale between 2.5 and 3.5 avoids plastic skin artifacts.
Can Midjourney flags like --ar or --stylize be used in FLUX or DALL-E?
No. FLUX.1 and DALL-E 3 do not parse Midjourney CLI flags. Passing CLI flags into FLUX or DALL-E can result in the model printing the raw flag text onto the image. Use our Universal Prompt Transpiler to convert flags automatically.