You expected sharp eyelashes, fabric weave, dust on the floor. What you got looks like a smartphone JPEG from 2010: the subject and pose are right, but the image is soft, plasticky, and reads “low quality” at a glance.
Fastest fix (90 seconds): raise the generation resolution to 1024 or higher, push sampler steps to 30, and paste 3-4 detail-anchor words into the prompt. On Midjourney (V8.1, default since June 10 2026) just turn on HD in Settings to render native 2K with no upscaling step. That single change resolves most “soft output” complaints. If detail is still missing after that, work down the table below.
Which bucket are you in?
Run these checks in order. The first one that matches is almost always your real cause.
| Check | If true | Cause | Jump to |
|---|---|---|---|
Resolution < 1024 on either axis (SDXL/Flux) | Output is uniformly soft, even in the prompt’s main subject | Too few pixels | Step 1 |
Steps at 20-25 with Euler / DDIM | Fine textures (hair, pores, fabric) mushy but composition fine | Under-denoised | Step 2 |
| Prompt has no detail words | Image is “fine” but flat, no standout detail | No detail anchors | Step 3 |
| Detail jumps when you swap checkpoints | Base model is the ceiling | Wrong checkpoint | Step 5 |
| CFG / guidance at the floor | Prompt loosely followed, anchors ignored | CFG too low | Common causes #5 |
| Looks fine in-app, bad after download | Only the exported file is soft | JPEG compression | Step 6 |
Common causes
Ordered by hit rate, highest first.
1. Generation resolution is too low
Many older defaults still use 512×512 or 768×768. That was fine for SD 1.5, but SDXL, Flux, and Midjourney V8 are trained on 1024×1024 and up, and they produce noticeably softer output below that. Below native resolution the model has fewer pixels to place detail into, so it averages textures into mush.
How to spot it: check your tool’s resolution setting. If either axis is < 1024 on SDXL / Flux / Midjourney, you are starving the model.
2. Sampler / step count too low
A default of 20 steps with a basic sampler under-resolves fine detail. Eyelashes, hair strands, and fabric texture all need more denoising iterations to converge. As of June 2026 the reliable SDXL sweet spot is DPM++ 2M Karras at 25-30 steps; the Karras schedule spreads noise reduction more evenly so detail holds up.
How to spot it: your step count is at 20-25 with Euler or DDIM. Move to DPM++ 2M Karras at 30 steps and re-run the same seed.
3. Prompt has no detail anchors
You wrote portrait of a woman. Nothing in that prompt says “detailed,” so the model generates its minimum aesthetic baseline, not a peak. Detail-anchor words pull it up.
How to spot it: scan the prompt for words like detailed, sharp, 8k, intricate, crisp, fine, texture, or lens specs. None present means soft output.
4. Wrong checkpoint for the subject
A general checkpoint produces general output. SDXL base is mid-detail on faces and weak on intricate fabrics; a realism-tuned checkpoint (Juggernaut XL, RealVisXL) produces much sharper results on the same prompt.
How to spot it: keep the same prompt, seed, and settings, then swap to a realism checkpoint. If detail jumps, the base model was the bottleneck.
5. CFG too low
Classifier-Free Guidance below 4 on SDXL, or guidance below 2.5 on Flux, makes the model follow the prompt loosely, including the detail anchors. Detail words get half-ignored.
How to spot it: CFG / guidance is at or below the default. Bump it 1-2 notches (SDXL 5-7, Flux 3-4) and re-test.
6. Output saved as low-quality JPEG
Your platform may auto-export a 70-quality JPEG for sharing. The generated pixels were fine; the export ruined them. This is the one cause where the in-app preview looks sharp but the downloaded file does not.
How to spot it: zoom into the downloaded file. If you see blocky compression artifacts around edges that were crisp in the app, it is the export, not the generation.
Shortest path to fix
Step 1: Raise the output resolution
# Midjourney (V8.1, default since 2026-06-10)
- Turn on HD: Settings panel -> Version section -> HD,
or append --hd to the prompt. HD renders native 2K (2x size, 4x
resolution of V7) with no separate upscale step.
- --hd costs ~1.3 GPU-min vs ~0.8 for SD; HD renders take noticeably
longer than SD for the extra pixels.
- Pick an aspect ratio with --ar 1:1 / 4:5 / 9:16; MJ sizes each axis to 1024+.
# Stable Diffusion / SDXL via Forge / A1111
- Width x Height: 1024x1024 minimum
- Enable "Hires. fix" -> upscale by 2x with R-ESRGAN 4x+
- Denoising strength: 0.35-0.5
# Flux dev (ComfyUI)
- 1024x1024 base
- Optional: chain a 2x upscale node
# ChatGPT (GPT Image 2.0, gpt-image-2, default since 2026-04-21)
- DALL-E 3 was retired from ChatGPT in Dec 2025; the in-app generator
is now GPT Image. Native output goes up to 2K.
- Ask explicitly: "Generate a high-resolution 1536x1024 image of ..."
- Plus/Pro users: use Thinking mode for self-verification and batching.Note: Midjourney’s --quality (alias --q) only accepts 0.25, 0.5, and 1 (default 1), and it changes detail/processing time, not pixel resolution. For more pixels you need HD mode or an upscale, not --quality.
Step 2: Increase sampler steps and switch sampler
# A1111 / Forge / SDXL
- Steps: 30 (was 20); 25-30 is the DPM++ 2M Karras sweet spot
- Sampler: DPM++ 2M Karras OR DPM++ SDE Karras
- CFG: 5-7
# ComfyUI Flux
- Steps: 28-35 (Flux needs fewer steps than SDXL)
- Sampler: euler + simple/normal scheduler
(Karras has minimal effect on Flux; do not force it)
- Guidance: 3-4
# Midjourney
- Leave --quality at 1 (the default). On older subscriptions it could
default to 0.5, so set --quality 1 explicitly if detail looks halved.Step 3: Add concrete detail anchors to the prompt
Paste these at the end. Pick the relevant set; do not stack all of them.
# For portraits
intricate skin texture, fine pore detail, individual eyelashes,
crisp focus on the eyes, 50mm lens, f/2.8 portrait,
shot on Hasselblad H6D
# For products / objects
hyperdetailed, fabric weave visible, micro-texture,
macro photography, focus stacking, 100mm macro lens
# For environments
dust particles in light beams, weathered surface texture,
fine architectural details, sharp foreground bokeh background
# Generic boost
hyperdetailed, intricate, sharp focus, photorealistic, 8k,
masterpiece, professional photography, ultra-detailedStep 4: Run a dedicated upscale pass
After base generation (skip this if you already used Midjourney HD, which is already 2K):
# Built-in upscalers
- Midjourney: "Upscale (Subtle)" keeps the image faithful;
"Upscale (Creative)" adds invented detail (use only when you want it)
- SDXL Forge: Extras tab -> R-ESRGAN 4x+
- ComfyUI: UltimateSDUpscale (USDU) custom node
# External upscalers (best quality)
- Topaz Gigapixel / Topaz Photo AI (paid subscription; best for portraits)
- Upscayl (free, open-source, Real-ESRGAN models; great for general use)
- Real-ESRGAN x4+ via ReplicateStep 5: Switch to a realism-tuned checkpoint
If the base model is the bottleneck:
# SDXL realism
- Juggernaut XL
- RealVisXL
- Realism Engine SDXL
- DreamShaper XL Turbo (faster)
# Flux realism
- Flux dev + Flux Realism LoRA (Civitai)
- Flux dev + Skin Realism LoRAStep 6: Save as PNG, not auto-JPEG
Always download the raw PNG, not the auto-compressed share preview. On Midjourney, open the image and use the download/save action for the full-quality file rather than copying the in-grid thumbnail.
How to confirm it’s fixed
- Zoom the new image to 100% (1:1 pixels). High-frequency areas (hair, pores, fabric weave, text edges) should stay crisp, not smeared.
- Compare against your old output at the same crop. The fix worked if texture detail is visibly present where it was mush before.
- Check the saved file, not the preview. Re-open the downloaded PNG and confirm there are no JPEG block artifacts around edges.
- If you want a controlled test, keep the same seed and change only one variable at a time (resolution, then steps, then prompt) so you know which lever actually moved the needle.
Prevention
- Standard hero-image settings: 1024+ resolution (or Midjourney HD), 30 steps, CFG 5-7, realism checkpoint, detail anchors in the prompt.
- Run an upscale pass before exporting; never ship a raw 1024 image to a 2K+ display.
- Save the raw PNG; never use an auto-shared JPEG preview as the final asset.
- Keep one saved preset / config per model that hits these settings by default, so you don’t re-debug this every session.
FAQ
Why does my image look sharp in the app but blurry after I download it? That is JPEG export, not generation. The share/preview file is re-compressed. Download the raw PNG (Step 6) and zoom in on the saved file to confirm.
Is higher --quality in Midjourney the same as higher resolution?
No. --quality only accepts 0.25, 0.5, and 1, and it changes detail and GPU time, not pixel count. For a bigger, sharper file use HD mode (or --hd) in V8.1, which renders native 2K.
More steps always means more detail, right?
Up to a point. SDXL with DPM++ 2M Karras converges around 25-30 steps; beyond that you mostly burn GPU time for tiny gains. Flux needs fewer steps still (around 28-35 with Euler). Past those ranges, fix the prompt or checkpoint instead of adding steps.
ChatGPT used to use DALL-E 3 and I can’t find those settings anymore. What changed?
DALL-E 3 was removed from ChatGPT in December 2025 and deprecated in the API on May 12, 2026. The in-app generator is now GPT Image (GPT Image 2.0 / gpt-image-2 as of April 21, 2026), which generates up to 2K natively. Ask for a specific size like 1536x1024 and a “high-resolution” image directly in your prompt.
My faces look soft no matter what I do. What’s the single biggest lever? Checkpoint. SDXL base is only mid-detail on faces. Swap to a realism checkpoint (Juggernaut XL or RealVisXL) at the same seed and settings; that one change usually does more for face detail than steps or resolution.