You generate a hero shot of your product. The lighting reads premium, the background reads agency-quality, but the bottle has a curve where there should be a straight edge, the box has a wobble in the side panel, and the logo on the front melted into a blurry approximation. For a real, branded product, no text prompt will fix this: the model has never seen your packaging, so it invents a plausible-looking fake.
Fastest reliable fix: stop generating the product from text. Either lock the silhouette with a ControlNet pass (Canny or Depth) from a real product photo in ComfyUI / Automatic1111, or skip generation entirely and composite the real cut-out product onto an AI-generated background. For commercial work, the composite route is faster and more accurate than fighting any model to perfection.
Common causes
Ordered by how often each is the actual root cause.
| # | Cause | Tell-tale sign | Primary fix |
|---|---|---|---|
| 1 | Model never saw your product | Every re-roll gives a different fake silhouette/label | ControlNet or composite (Steps 1, 4) |
| 2 | No structural constraint | Silhouette changes seed to seed | ControlNet Canny/Depth (Step 1) |
| 3 | Style fighting realism | Heavy art-style modifier warps the shape | Strip style from product (Step 3) |
| 4 | Too few pixels on the product | Label/logo soft only when product is small in frame | Reframe for pixel budget (Step 2) |
| 5 | Off-distribution camera angle | Distortion at top-down or extreme angles only | Use straight-on / three-quarter |
| 6 | Multiple products in one frame | Each unit slightly wrong in group shots | Composite each unit (Step 4) |
1. The model has never seen your product
Diffusion models know “bottle”, “box”, “phone” as abstract categories. They cannot reproduce your specific silhouette, embossing, or label layout from text alone. Every text-only render of a branded product is a hallucination of plausible packaging. This is the root cause in the large majority of cases.
How to spot it: Are you generating a real, branded product purely from a text prompt? If yes, this is your problem and prompt engineering will not fix it.
2. No structural constraint on the silhouette
The model picks a generic shape that fits “bottle” or “box” and improvises. Without a ControlNet (Canny, Depth) or a reference image, every render has a different silhouette.
3. Stylization fighting realism
product photo of a Coke bottle in the style of Andy Warhol pulls the bottle toward the Warhol aesthetic, and the silhouette degrades. Style modifiers and accurate product geometry compete for the same generation budget.
4. Too few pixels on the product region
Embossed logos, fine type, and thin features (handles, spouts, caps) need pixels. A 1024x1024 hero of a tall thin bottle gives the bottle maybe 200px of width, not enough for label fidelity. Larger base resolutions (FLUX.2 and Midjourney V8.1 render native 2K) help, but reframing helps more.
5. Off-distribution camera angle
Models trained on catalog shots handle straight-on and slight three-quarter views well. Top-down, worm’s-eye, or extreme angles push into low-data territory and shapes distort.
6. Multiple products in one frame
Group shots (a six-pack, a starter set) ask the model to render every unit correctly. Even with constraints, each unit competes for accuracy budget, and the failure rate multiplies.
Before you start
- Decide whether the image needs the real product or whether a “product-like” stand-in is acceptable. Concept work tolerates stand-ins; commercial work does not.
- Save the seed, prompt, model, and tier of the broken image so you can iterate from the same baseline.
- Have at least one clean reference photo of the real product on a neutral background.
- Confirm the use case. Print needs more accuracy than web; packaging mockups need pixel-perfect labels.
Information to collect
- Full prompt, model, seed, sampler, steps, aspect ratio.
- A reference photo of the real product, ideally on a clean background.
- The features that must read correctly (logo position, label color, silhouette curve).
- Intended deliverable size and use (web hero, print poster, social card).
Step-by-step fix
Ordered by ROI. Step 1 plus Step 4 is the standard production workflow for commercial hero work.
Step 1: Lock the silhouette with ControlNet
The single biggest move is replacing text-only generation with a structural constraint taken from the real product. ControlNet lives in the open-weight stack (Automatic1111, ComfyUI), not in hosted apps like Midjourney.
- SDXL (A1111 / ComfyUI): Load a Canny ControlNet from a product photo with a clean silhouette. Start at control weight
1.0for strict adherence; drop toward0.7if edges look traced/flat. - Rounded surfaces (bottles, jars): Use Depth ControlNet instead of Canny. Depth follows curvature where Canny only follows hard edges.
- FLUX.2 (ComfyUI): Use the Flux ControlNet Union (Shakker-Labs
FLUX.1-dev-ControlNet-Union-Pro-2.0, or the FLUX.2 Fun Controlnet Union for FLUX.2-dev). One model covers Canny, Depth, Soft Edge, Pose, and Grayscale, and matches the exact object boundary of your input. Preprocess withComfyUI ControlNet aux. - Midjourney (V8.1, default since June 10 2026): Midjourney has no ControlNet. The closest tools are
--sref+--sw(style) and image prompts. As of June 2026 these do not reliably lock a branded product’s geometry, so for Midjourney route product accuracy through composite (Step 4) instead.
ControlNet locks the silhouette; the model fills in lighting, environment, and surface. The shape stays yours.
Step 2: Raise the pixel budget for the product region
Pick framing that gives the product more pixels:
- Tall product -> vertical aspect ratio.
- Wide product -> horizontal.
- Square product -> square.
- Crop tighter to the product if the environment is not the story.
Then upscale the final image rather than generating large from the start; product geometry is more stable at the model’s native resolution.
Step 3: Remove style fights from the prompt
Drop heavy art-style modifiers. Lead with photographic descriptors:
product photography, studio lighting, sharp focus, clean background,
[product description], shot on HasselbladPush style modifiers onto the background and lighting, never onto the product surface.
Step 4: Composite the real product photo
This is the standard production workflow for commercial hero work, and the only fully reliable path for Midjourney-style hosted tools:
- Generate the AI image with a generic stand-in (a “blank white bottle” or “blank box”) so the lighting and scene are right.
- Export the AI background and lighting.
- Cut the real product out of its reference photo (Photoshop
Select Subject/Remove Background, orrembg). - Place the cut product into the AI background.
- Match lighting direction with a soft drop shadow, then apply one global color grade over the whole frame so the composite reads unified.
For most commercial hero work this is faster and more accurate than tuning ControlNet to perfection.
Step 5: Mask and inpaint the product region
If the composition is locked but the product shape is still slightly off:
- Mask the product area.
- Run img2img inpaint with the ControlNet still active.
- Reduce denoise to
0.3-0.4to preserve the silhouette while refining surface detail.
Step 6: Switch to a stronger model
As of June 2026, for product silhouettes and labels:
- FLUX.2 Pro and Midjourney V8.1 hold product geometry and short label text far better than SDXL base or any SD 1.5 model.
- Ideogram v3 (Quality) is strongest for readable label/logo text (independent tests put it near 90% text accuracy, ahead of the field).
- Imagen 4 is the most reliable for correctly spelled words on signs and labels.
Route the silhouette to ControlNet/FLUX, and route any in-image text to Ideogram v3 or composite it from the real artwork.
How to confirm the fix
- Overlay the generated product on the reference photo at the same scale. Silhouette edges should align within a few pixels.
- Inspect the logo and label region at 100% zoom. Letters should be readable and correctly spelled.
- Check straight edges on boxes: they should be straight, not wobbled.
- Check axisymmetric features on bottles and jars: the two sides of the silhouette should mirror.
- For multi-unit shots, every unit must pass all of the above.
Long-term prevention
- For any commercial product work, default to ControlNet plus composite. Do not rely on text-only generation.
- Maintain a clean reference-photo library for every product you make hero shots of.
- Use FLUX.2 Pro or Midjourney V8.1 for product work. Avoid SD 1.5 entirely.
- Keep a saved product-photography prompt template that excludes heavy style modifiers from the product itself.
- For labels and logos, composite type from the real product image or vector file. Never trust the model with brand-critical text.
Common pitfalls
- Re-rolling 20 seeds hoping the model nails your product. It will not.
- Adding
accurate product shapeto the prompt expecting magic. It does almost nothing. - Compositing the product without matching lighting direction, so the composite reads fake.
- Forgetting to inspect the label at 100% zoom. Label errors hide at thumbnail size.
FAQ
Q: Can the model learn my product if I train a LoRA? A: Yes, for SDXL and FLUX. A product LoRA trained on 20-50 photos holds silhouette and label fidelity well. It is worth the setup for products you generate often, and it stacks with ControlNet.
Q: Does ControlNet Canny work for transparent products like glass bottles? A: Only partially. Canny detects edges, and transparent products have weak edges. Use Depth ControlNet (or train a LoRA) for glass and jars.
Q: Which model gives the most accurate product labels in 2026? A: Ideogram v3 (Quality) for readable label text, and Imagen 4 for correctly spelled words. For brand-critical work, do not generate the label at all: composite it from the real artwork.
Q: How do I do this in Midjourney if it has no ControlNet?
A: You cannot lock a branded silhouette in Midjourney directly. Use it to generate the scene and lighting with a generic stand-in, then composite the real product on top (Step 4). --sref controls style, not product geometry.
Q: What about generating packaging mockups from scratch? A: Use a dedicated mockup workflow (Smart Mockups, Placeit, or a Photoshop smart-object template) and overlay your AI-generated label. Cleaner than asking the model to invent packaging.
Q: Why does my logo come out almost right but slightly off? A: The model rasterizes the logo as an image feature it can only approximate. For brand-critical work, composite the real logo from a vector file on top of the AI render.