Juggernaut XL Photorealism: Settings, Prompts & Fixes

Last updated: August 28, 2026

7-Step Settings Checklist for Fast Photorealistic Results

Follow each step in order before you generate your first image. Skipping steps usually causes plastic skin and unstable faces at scale.

1. Set resolution to 832×1216 (portrait) or 1216×832 (landscape). These SDXL-native dimensions work for both v9 and Ragnarok. Running at 1024×1024 or random sizes often breaks facial coherence.
2. Select the model-specific sampler. Use DPM++ 2M Karras for v9 and DPM++ 2M SDE for Ragnarok. The v9 model card specifies DPM++ 2M Karras, and Ragnarok’s card specifies DPM++ 2M SDE. See the comparison table below for the full settings snapshot.
3. Set steps to 30–40. Both model cards recommend this range. Lower step counts tend to blur skin transitions.
4. Set CFG between 3 and 5 for photorealism. Ragnarok recommends CFG values in the 3–6 range, with lower values producing more realistic results. Values above 7 usually create waxy, plastic skin that fails commercial use.
5. Skip the negative prompt on your first run. The official v9 card suggests starting with no negative prompt. Add only the specific issues you actually see.
6. Enable Hi-Res Fix with 4xNMKD-Siax_200k, 15 steps, 0.3 denoise, 1.5× scale. These upscaler values come directly from the v9 model card. Ragnarok widens denoise slightly to 0.3–0.33 to reveal more skin pores.
7. Use the baked-in VAE for v9. The v9 model card confirms the VAE is baked in. Loading a second VAE usually breaks color response.

Sozee’s Photo Control and Photo Shoot system mirrors these direction choices as setting, outfit, shot style, expression, and object. You never touch a sampler field. Get started and produce locked-likeness commercial assets today.

Side-by-Side Settings for Juggernaut XL v9 and Ragnarok

Juggernaut XL v9 integrates RunDiffusion Photo v2, which improves skin detail, micro-texture, and lighting control compared to v8. Lower CFG keeps the diffusion process close to the data manifold instead of over-amplifying classifier gradients. Skin then reads as photographed instead of rendered. Every CFG step above 5 nudges results toward an illustrated, plastic look that often fails paid campaigns.

How Ragnarok Pushes Juggernaut Photorealism Further

Juggernaut XL Ragnarok (version XIII) launched in July 2026 as the final SDXL model in the series. It starts from version 12 as the photoreal base and merges two NSFW SDXL fine-tunes at low ratios of 0.15 and 0.1. This training recipe keeps skin pores and lighting falloff as core strengths. Ragnarok favors studio portrait lighting with clean contours and soft skin transitions compared to NSFW-first models.

Ragnarok works best with CFG between 3 and 6, and lower values in that band usually look more realistic. This lower CFG sweet spot pairs with the 0.3–0.33 Hi-Res Fix denoise range, which resolves pores and lighting falloff at the final size without over-smoothing. These photoreal gains extend to anatomy, so hands and full-body poses look more natural than in earlier versions.

Photographic Prompt Formula That Juggernaut Understands

Use this order for every Juggernaut prompt: subject → lighting → camera → film stock → post-processing. Juggernaut XL responds best to this structured positive sequence. Also, SDXL models tend to ignore content after about 75 tokens, so place subject and lighting first, then camera and quality tags.

Five ready-to-use examples:

1. RAW photo, cinematic mid shot of a 28-year-old woman with dark brown hair, soft window light from camera left, 85mm f/1.4, Kodak Portra 400, shallow depth of field, skin details, hyperdetailed photography
2. RAW photo, full body photo of a woman in a white linen dress standing in a sunlit courtyard, golden hour backlight, 50mm f/2.0, film grain, natural skin texture, pores visible, 8k uhd
3. Cinematic still mid shot photo of a woman at a café table, overcast diffused daylight, 35mm f/2.8, Fuji 400H color science, hyperdetailed photography, skin details
4. RAW photo, close-up portrait of a woman with freckles, candlelight from below, 105mm f/1.8 macro, warm color grade, detailed skin texture, cinematic movie
5. RAW photo, three-quarter shot of a woman in a leather jacket on a rooftop at dusk, neon ambient fill, 85mm f/1.8, Cinestill 800T, shallow depth of field, skin details, sharp focus

Camera-specific language like “Canon EOS R5, 85mm f/1.8, golden hour lighting, detailed skin pores, film grain” consistently beats generic phrases such as “realistic photo of a person.”

Negative Prompt Strategy That Actually Helps

The official v9 model card suggests starting with no negative prompt. Heavy negative blocks often hurt more than they help. A practical workflow looks like this. Generate with an empty negative prompt, review repeated issues across several images, translate each issue into a clear visual noun, add the smallest set of terms, then remove any term that does not visibly improve results.

Here is a minimal working negative prompt for commercial portraits, built from real failure modes:

• Skin artifacts: plastic skin, waxy skin, airbrushed, over-smoothed, doll-like
• Face errors: asymmetrical eyes, dead eyes, uncanny valley, distorted face
• Hand and anatomy artifacts: extra fingers, extra limbs, deformed hands, fused fingers
• Realism drift: cartoon, painting, illustration, 3D render, CGI, digital art
• Polish junk: watermark, text, signature

Avoid pasting a 200-token negative prompt on your first run. When skin looks plasticky or shiny, first lower CFG in 0.5 steps toward 3. Skin issues usually start with CFG, not with the prompt.

ComfyUI Workflow for Batch-Scale Juggernaut Sessions

A minimal ComfyUI graph for Juggernaut photorealism at batch scale uses this node chain: Load Checkpoint, CLIP Text Encode (positive and negative), KSampler, VAE Decode, Save Image. For sessions with more than 50 assets, extend this graph with IP-Adapter and FaceDetailer nodes to hold identity across the batch.

Use these node settings for consistent batches:

• KSampler: sampler_name = dpmpp_2m, scheduler = karras (v9) or dpmpp_2m_sde / karras (Ragnarok), steps = 35, cfg = 3.5, denoise = 1.0
• IP-Adapter-Plus-Face (ip-adapter-plus-face_sdxl_vit-h): weight = 0.6–0.8, noise = 0.0, End At = 0.8, which keeps the background flexible while the prompt refines details later in sampling
• Secondary IP-Adapter (style or clothing): weight = 0.4 on a separate stream from the face adapter to avoid style bleed
• FaceDetailer (Impact Pack): guide_size = 256, denoise = 0.3–0.4, which re-projects IP-Adapter influence onto the face at high resolution and fixes uncanny glitches without full regeneration
• VAE Decode Tiled: replace standard VAE Decode with VAE Decode Tiled for large batches to avoid VRAM exhaustion during decoding

A1111 vs ComfyUI for 50+ asset sessions: A1111’s X/Y/Z plot handles sampler and CFG sweeps but needs manual seed control and lacks built-in IP-Adapter face locking. ComfyUI’s node graph runs face reference injection, FaceDetailer, and tiled decode in one unattended pipeline. A saved ComfyUI workflow JSON can run overnight and generate hundreds of consistent variants, with quality enforced by the graph instead of per-run tweaks. For commercial likeness locking across many clients, ComfyUI is the only practical local choice.

Sozee replaces this entire setup. Upload three reference photos, set five direction dimensions, and Photo Shoot returns a locked, coherent set of up to ten images per frame with the same face and body in every output. Start creating now and skip the node graph entirely.

Fixing Skin, Lighting, and Likeness Problems

Plastic or waxy skin: This issue almost always traces back to CFG above 5. Lower CFG in 0.5 steps until the wax-figure effect disappears before you change the negative prompt. If skin still looks synthetic at CFG 3, add waxy skin, plastic skin, doll-like to the negative prompt as focused terms instead of a long quality list.

Lighting falloff loss: Flat lighting usually comes from the prompt, not the sampler. Juggernaut XL reacts strongly to directional lighting prompts such as golden hour, window light, or candlelight. These phrases shape skin tone and reduce a synthetic look. Add a clear light direction and color temperature to the positive prompt before you touch any other setting.

Likeness drift across a batch: In ComfyUI, if characters start to look doll-like or lose identity, slightly lower the IP-Adapter weight or strengthen the text prompt with traits such as “freckles, sharp jawline, messy brown hair.” This balance between visual reference and text tokens restores a stable likeness. In A1111, batches larger than about ten images drift without IP-Adapter support. Perfect consistency across 50 or more images remains difficult in 2026 because drift accumulates in faces, clothing, and proportions.

Frequently Asked Questions

How do Juggernaut XL v9 and Juggernaut XIII Ragnarok differ for photorealism?

Juggernaut XL v9 serves as the established SDXL photorealism workhorse with a wide tooling ecosystem and a baked-in VAE. It uses DPM++ 2M Karras, CFG 3–7, and 30–40 steps at 832×1216. Juggernaut XIII Ragnarok is the final SDXL evolution, released in July 2026, with stronger anatomy, better hands, and a photoreal-first base merged with NSFW fine-tunes at low ratios. Ragnarok uses DPM++ 2M SDE, recommends CFG 3–6 with lower values for realism, and extends Hi-Res Fix denoise to 0.3–0.33 for extra pore detail. Choose Ragnarok for commercial portrait work that needs the anatomy upgrade. Choose v9 when you want maximum ecosystem compatibility and plugin support.

Why does my Juggernaut output still look plastic?

Plastic skin on Juggernaut almost always comes from CFG. Values above 5–6 push classifier gradients so far that skin stops reading as photographed. Lower CFG in 0.5 steps toward 3 before you change anything else. If plastic skin remains at CFG 3, add focused negative terms such as waxy skin, plastic skin, doll-like instead of a long quality block. Heavy negative prompts often remove detail that Juggernaut would otherwise handle well. Also confirm Hi-Res Fix denoise sits between 0.3 and 0.33. Values above 0.4 can re-render skin in a way that looks synthetic at the final resolution.

What negative prompt works best for Juggernaut photorealism?

Start with an empty negative prompt on your first run. Review repeated issues across several images, then add only the terms that match those issues. A lean, effective set for commercial portraits covers five areas. Skin artifacts include plastic skin, waxy skin, airbrushed, and doll-like. Face errors include asymmetrical eyes, dead eyes, and distorted face. Anatomy failures include extra fingers, extra limbs, and fused fingers. Realism drift includes cartoon, painting, 3D render, and CGI. Polish junk includes watermark, text, and signature. Avoid 200-token community negatives, because they usually suppress useful detail.

How can I keep character likeness stable across 50 or more Juggernaut images?

Local Juggernaut workflows need a ComfyUI pipeline with dual IP-Adapter. Use one IP-Adapter-Plus-Face stream at weight 0.6–0.8 for identity, a second IP-Adapter stream at weight 0.4 for clothing or style, and a FaceDetailer pass at denoise 0.3–0.4 to fix uncanny drift without full regeneration. This setup holds likeness across moderate batch sizes, although SDXL still drifts beyond about 50 images. Sozee’s Photo Shoot system instead locks likeness from three reference photos and keeps the same face, body, and world across every frame in a set, with no node graphs, sampler tuning, or batch rerolls.

Conclusion: Scale Photorealistic Assets Without the Heavy Workflow

Juggernaut XL v9 and XIII Ragnarok can both deliver photorealistic images that pass as professional photography. Achieving that level consistently across 50 or more assets for paying clients demands careful sampler choice, low-end CFG discipline, a structured prompt formula, a lean negative prompt based on observed issues, and a ComfyUI pipeline with IP-Adapter and FaceDetailer. Each layer adds both quality and complexity.

Sozee removes that complexity. Upload three photos, set five direction dimensions in Photo Control, and Photo Shoot returns a locked, coherent set where face, body, and environment stay consistent across every frame. No sampler menus, no node graphs, no endless rerolls. You reach the same commercial output in an afternoon instead of a week.

Go viral today, build your locked-likeness character, and start producing commercial assets at scale.