G858-debug/docs/skills/juggernaut-ragnarok

Juggernaut XL Ragnarok — Pipeline Reference

Overview

Juggernaut XL Ragnarok is a photorealistic SDXL checkpoint. It is the most downloaded SDXL model (520K+ downloads) and the final SDXL release from KandooAI / RunDiffusion.

Key characteristics:

• Photorealistic output with cinematic quality
• NSFW capability baked into training (trained with Booru tags on an NSFW dataset, merged with a Lustify-based NSFW pass for anatomical stability)
• Supports BOTH natural language prompts AND Booru-style tags
• No safety filter — ComfyUI has no built-in NSFW filter; content is controlled entirely through prompts
• Uses natural language for SFW scenes and Booru tags for NSFW anatomical detail

Checkpoint file: Juggernaut-Ragnarok.safetensors Architecture: SDXL 1.0 VAE: Baked in (no separate VAE needed) License: Check Civitai/HuggingFace for current license terms

Generation Settings

| Parameter | Value | Notes | |-----------|-------|-------| | Resolution | 832×1216 (portrait), 1024×1024 (square) | Use standard SDXL resolutions | | Sampler | DPM++ 2M SDE | Primary recommendation | | Scheduler | Karras | | | Steps | 30–40 | 30 for speed, 40 for quality | | CFG | 3–5 | Lower = more realistic. 3 for maximum realism, 5 for more prompt adherence | | Denoise | 1.0 | For txt2img. Use 0.2–0.4 for img2img refinement |

HiRes / Upscale Settings

| Parameter | Value | |-----------|-------| | Upscaler | 4xNMKD-Siax_200k | | Upscale Steps | 15 | | Upscale Denoise | 0.3 | | Upscale Factor | 1.5x |

Prompt Structure

Component Order (earlier = more weight)

Ragnarok respects prompt token order — place the most important elements first.

1. Subject — The primary focus (woman, man, two people, etc.)
2. Action — What the subject is doing
3. Clothing — CRITICAL for SFW. Ragnarok defaults toward nudity due to NSFW training. Always describe clothing explicitly for SFW images.
4. Expression/Emotion — Facial expression, eye contact direction
5. Environment/Setting — Where the scene takes place
6. Lighting — Specific light source (never generic)
7. Style/Medium — "photograph", "cinematic", "photorealistic"
8. Perspective/Composition — Camera angle, shot type, depth of field
9. Color/Atmosphere — Mood, color palette, atmospheric elements
10. Texture/Material — Surface details, fabric descriptions

Prompt Size

Try not to exceed 75 tokens. SDXL has a 77-token CLIP limit — content beyond this is truncated. Be concise and specific rather than verbose.

Quality Boosters (optional)

For maximum photorealism, prepend:

masterpiece, 4k, ray tracing, intricate details, highly-detailed, hyper-realistic, 8k RAW Editorial Photo

For film/analog aesthetic:

film grain ISO 200 faded film, 35mm photo, grainy, vignette, vintage, Kodachrome, Lomography, stained, highly detailed, found footage

For clean photographic look:

photograph, high resolution, cinematic, skin textures, detailed skin

SFW Prompting Rules

CRITICAL: Because Ragnarok was trained on NSFW data, SFW images require explicit clothing descriptions. Without clothing in the prompt, the model may default to nudity or revealing states.

Always include in SFW prompts:

• Specific clothing descriptions (e.g., "elegant dress", "fitted blazer and tailored trousers", "casual streetwear")
• Add NSFW-prevention tokens to negative prompt: nudity, naked, nsfw, topless, nude

SFW negative prompt template:

nudity, naked, nsfw, topless, nude, bad eyes, blurry, missing limbs, bad anatomy, cartoon, low quality, worst quality, deformed

NSFW Prompting Rules

For explicit/intimate content, use Booru-style tags for anatomical precision:

• Booru tags give more reliable anatomical results than natural language for NSFW
• Keep the scene description in natural language, use Booru for body/positioning
• The model was specifically trained with Booru-tagged NSFW data

NSFW prompt pattern:

[Natural language scene description], [Booru body/position tags], [lighting], [composition]. Photorealistic.

Standard negative prompt (NSFW):

bad anatomy, bad hands, extra limbs, watermark, blurry, text, cartoon, illustration, painting, low quality, worst quality, deformed

What NOT to Do

• Don't use Pony quality tags — score_9, score_8_up, source_pony etc. are Pony-specific and meaningless to Ragnarok
• Don't use rating_safe / rating_explicit — these are Pony rating tags
• Don't leave clothing out of SFW prompts — the model will default toward nudity
• Don't use CFG above 7 — causes waxy skin, oversaturation, and unrealistic output
• Don't exceed 75 tokens — content is truncated beyond the CLIP limit
• Don't use generic lighting — "warm lighting" is vague. Name the source: "single amber streetlight", "candlelight", "golden hour through window"

LoRA Compatibility

• Ragnarok is SDXL 1.0 architecture — compatible with all SDXL LoRAs
• Character LoRAs should be trained against SDXL 1.0 base (not Ragnarok itself) for maximum portability across SDXL fine-tunes
• LoRA injection via ComfyUI LoraLoader nodes — same pattern as Pony pipeline
• Keep total LoRA count minimal (1-2 per generation) to preserve prompt adherence. The Pony pipeline used 6+ style LoRAs — this is NOT recommended for Ragnarok. The checkpoint handles photorealism natively.

Skin Tone and Ethnicity

Ragnarok uses natural language for ethnicity and skin tone — not Booru tags.

For Black South African characters across the full skin tone spectrum:

• Describe skin tone explicitly: "light brown skin", "medium-brown skin", "dark brown skin", "deep brown skin", "warm caramel skin"
• Include ethnicity naturally: "Black South African woman", "young Zulu man"
• Hair texture and style should be described independently of skin tone
• African identity is expressed through the combination of facial features + hair texture + skin tone — not skin darkness alone

Testing guidance: Before committing to production, generate test batches at various skin tones and lighting conditions to verify the checkpoint renders them well without supplementary LoRAs.

Comparison with Previous Pipeline (CyberRealistic Pony)

| Aspect | CyberRealistic Pony | Juggernaut Ragnarok | |--------|---------------------|---------------------| | Aesthetic | Semi-realistic | Photorealistic | | Prompt style | Booru tags only | Natural language + Booru for NSFW | | Quality tags | score_9, score_8_up | masterpiece, 4k, hyper-realistic or none | | CFG | 5.0 | 3–5 | | Style LoRAs needed | 6 (Ebony Pony, Skin Tone, etc.) | 0 — checkpoint handles photorealism natively | | NSFW capability | Via rating tags | Baked in — clothing required for SFW | | Negative prompt style | Pony-specific negatives | Standard SDXL negatives | | Sampler | dpmpp_2m_sde / karras | dpmpp_2m_sde / karras (same) |