The image is clean, well-lit, and technically pretty, but it looks like a frame from a 3D animated movie rather than a photograph. Skin reads like wax, hair looks like polygons, and the lighting is too perfect. Your brain calls “CGI” in under a second.
Fastest fix (works most of the time): delete every render keyword (octane render, unreal engine, 8k, hyperrealistic) from the prompt, drop CFG to 5-6 on SDXL or guidance to 2.5-3.5 on Flux, and switch to a photo-tuned model (Juggernaut XL, RealVisXL V5, or Flux.1 Krea). That combination alone moves most images from “render” to “photo.”
This is one of the most common SDXL / Flux symptoms, and it has five independent causes that stack. Fix one and it improves; fix the top three and it usually becomes photographic.
Which bucket are you in?
Run this quick triage before changing anything. The cause is almost always in the top rows.
| Check | If true, your cause is | Jump to |
|---|---|---|
Prompt contains octane render, unreal engine, blender, vray, cinema 4d, ray tracing | Render keyword contamination | Step 1 |
Prompt contains hyperrealistic, ultra detailed, 8k, 4k uhd, masterpiece | ”Quality booster” words drifting toward render | Step 1 |
SDXL CFG > 7, or Flux guidance > 4 | Guidance too high (plastic, oversaturated) | Step 3 |
| Same prompt looks photo on Juggernaut/RealVis but CGI on your model | Wrong checkpoint | Step 2 |
Lighting words are studio lighting, perfect lighting, three-point | Over-controlled CG-style lighting | Step 1 |
| Everything above is clean and it is still too smooth | Missing imperfection / grain | Steps 4-6 |
Common causes
Ordered by hit rate, highest first.
1. Render keywords contaminating the prompt
"portrait photograph, octane render, unreal engine 5, 8k, hyperrealistic, ray tracing"You said photograph, but you also said octane render and unreal engine. Those phrases dominate because they appear constantly in 3D-render training captions. The model averages between photo and CGI, and leans CGI.
How to spot it: scan the prompt for octane render, unreal engine, blender, vray, 3d render, cinema 4d, ray tracing, pbr, subsurface scattering. If any of these are present in a photo prompt, they are the problem.
2. Wrong checkpoint / model
Some popular models bake in CGI tendencies: Cinematic Diffusion, certain “photoreal” merges, and the fast distilled checkpoints (Flux schnell, Flux.2 [klein]) used at too few steps. They look amazing for stylized work but struggle with true-photo realism.
How to spot it: keep the same prompt and settings, swap to a known-photo checkpoint (Juggernaut XL, RealVisXL V5, or Flux.1 Krea). If the output becomes truly photographic, the model was the source.
3. CFG / guidance too high
When CFG (SDXL) or guidance (Flux) goes above the sweet spot, the model overcommits: colors get punchy, lighting gets dramatic, and surfaces go glossy and plasticky. That is the “overprocessed” CG look. On Flux dev this matters even more because guidance is distilled into the model, so a value of 3-4 already does roughly what CFG 7 does on SDXL. Push it higher and you get blown-out, artifact-heavy skin.
How to spot it: SDXL CFG > 7, Flux guidance > 4. Try CFG 5-6 (SDXL) or guidance 2.5-3.5 (Flux dev) and compare.
4. “Hyperrealistic” / “8k” pushing toward render
Counter-intuitive but well documented: words like hyperrealistic, ultra detailed, 8k, 4k uhd, and masterpiece are over-represented in 3D-render and CG-art captions. They drift output toward CGI, not photography. As of June 2026 the prompting consensus is the same: do not stack quality boosters to chase realism, because they make images more artificial, not less.
How to spot it: prompt has hyperrealistic, ultra-realistic, 8k, or cinematic 4k. Replace with photograph, 35mm film, documentary photography, or RAW photo.
5. Lighting prompted as “perfect” / “studio”
studio lighting, perfect lighting, three-point lighting, and ring light all push toward controlled, CG-like setups. Real photos have imperfections: mixed color temperature, color spill, and shadow noise.
How to spot it: the lighting words are clinical or perfect. Replace with available natural light, window light only, or motivated practical lamps.
Shortest path to fix
Step 1: Strip render keywords, add photo keywords
Remove from the prompt:
octane render, unreal engine, blender, vray, cinema 4d,
3d render, ray tracing, pbr shader, subsurface scattering,
cgi, computer graphics, hyperrealistic, 8k, masterpieceAdd photo-anchor words:
# General photographic
35mm film photograph, RAW photo, photojournalism,
documentary style, candid shot, photographed on Kodak Portra 400
# For portraits specifically
shot on Hasselblad H6D, 50mm f/1.4, available natural light,
unedited, soft skin imperfections visible
# Add slight imperfection (most powerful)
slight film grain, organic skin texture, natural color cast,
visible pores, fine flyaway hair, slight motion blur, imperfect focusNaming a real camera, lens, and film stock is more effective than abstract quality words. On Flux models in particular, Shot on Fujifilm X-T5, 35mm f/1.4 beats professional photo for authentic output.
Step 2: Switch to a true-photo checkpoint
Switch from any “stylized” or “cinematic” model to a realism-tuned one. Current strong picks as of June 2026:
# SDXL — strongest photo realism
- Juggernaut XL (v10 / "Ragnarok") — best all-round, character + grain
- RealVisXL V5.0 — tight portraits, natural film look
- CyberRealistic XL — cleanest output, but can feel sterile/editorial
# Flux — best photo realism
- Flux.1 Krea [dev] — "opinionated" model trained to kill the oversaturated AI look
- Flux dev + UltraReal Fine-Tune LoRA
- Flux dev + Skin Realism LoRA (Civitai)
- Turn OFF any anime / illustration LoRA still loadedAvoid for photo work: cinematic, fantasy-art, anime, and illustration-tuned merges. Flux.1 Krea was built specifically to avoid oversaturated, overly smooth textures, so if the CGI look is your main complaint it is often the single biggest fix.
Step 3: Drop CFG / guidance
# SDXL
CFG: 5-6 (was 8-10)
# Flux dev
Guidance: 2.5-3.5 (was 4-7)
# Midjourney V7
--stylize 50 (default is 100; lower = less artistic interpretation)
--style raw (strips Midjourney's house aesthetic — the bigger realism lever)Lower guidance means looser prompt adherence, which means more photographic naturalism. On Midjourney V8 / V8.1 the default already leans photographic, so reaching for --style raw plus a low --stylize is usually enough; the old “default 500” advice applied to V5/V6 and no longer matches the current default of 100.
Step 4: Add film / camera grit
Real photos have imperfections. Add one or two of these:
slight film grain, kodak portra 400 35mm film,
natural skin imperfections, subtle out-of-focus motion,
soft shadows with noise, ambient light bleedStep 5: Negative-prompt the CG signature (SD family)
3d render, cgi, octane render, unreal engine,
overprocessed, plastic skin, waxy skin, poreless skin,
airbrushed, oversaturated, hyperreal, smooth perfect skin,
smooth perfect hair, ray traced, sss, computer graphics, video gameCaveat by model: SDXL uses negative prompts natively. Plain Flux dev ignores a negative prompt unless your workflow uses a true-CFG node (such as a “Dynamic Thresholding” or CFG-guidance setup). Flux.2 does not support negative prompts at all, so on Flux.2 describe only what you want and lean on Steps 1-4 instead.
Step 6: Apply a film-grain LoRA / post-pass
If everything else is right but the image still looks too clean:
- SDXL: add
<lora:film_grain_v2:0.4>from Civitai - Flux: add a Flux Filmgrain LoRA at 0.3-0.5
- Post: pass the output through Filmkonvert, Dehancer, or a FilmStock plugin for analog character
- Quick post: open in any photo editor and add 10-15% film grain plus a slight color cast
How to confirm it’s fixed
Run an A/B in one batch: generate one image with your old prompt/settings and one with the changes from Steps 1-3, identical seed. The fixed version should have visible skin pores or texture, non-uniform lighting (one side warmer or in shadow), and slightly muted, non-neon color. If skin is still glassy and lighting is still flawless, you have not lowered guidance enough or the checkpoint is still a stylized one. Zoom to 100% on a cheek or forehead: real photos show micro-texture and grain there, renders show a smooth gradient.
Prevention
- Maintain two strictly separate prompt templates, photographic and 3D, and never blend them.
- Make low CFG / guidance the default for photo work; only raise it when prompt-following matters more than realism.
- Always include at least one imperfection word in photo prompts (grain, available light, candid).
- For series work, train a personal “photo-style” LoRA on real photo references; it locks the look far better than prompts.
FAQ
Why does adding “hyperrealistic” or “8k” make my image look worse?
Those tags appear far more often in 3D-render and CG-art captions than in real photo captions, so they pull the model toward render aesthetics. Use photograph, RAW photo, or 35mm film instead.
Lowering CFG made the image too vague and ignored my prompt. What now? That is the trade-off: lower guidance gives realism but looser adherence. Find the middle ground (SDXL CFG 6, Flux guidance 3.5), and put the must-have details earlier in the prompt so they survive at lower guidance.
Negative prompts don’t seem to do anything on Flux. Am I doing it wrong? Probably not. Plain Flux dev ignores negatives unless your workflow adds a true-CFG node, and Flux.2 dropped negative-prompt support entirely. On Flux, fix the look with positive imperfection words and a lower guidance value instead.
Which single change helps most? Switching to a realism-tuned model. On Flux that is Flux.1 Krea [dev], which was trained specifically to remove the oversaturated, plastic “AI look.” On SDXL it is Juggernaut XL v10 or RealVisXL V5.
My faces are great but the skin is still plastic-smooth. How do I fix just that?
Add visible pores, organic skin texture, fine flyaway hair, subtle wrinkles to the positive prompt, drop guidance slightly, and finish with a light film-grain pass (Step 6). Skin texture is usually the last 10% after the bigger fixes land.