Fabric has two classic AI failure modes, and a throw blanket exposes both. The first is texture: models tend to average a weave into a smooth, plastic-looking surface, because they reproduce the look of fabric without the structure of it. The second is pattern: a plaid or a stripe has to bend correctly as the cloth folds and drapes, and the documented failure is the pattern going wavy or losing its grid across the folds. A draped plaid wool throw puts both to the test at once.
So I ran it. One brief, a plaid wool throw in forest green and cream draped over a chair, visible weave and a fringed edge, through four of the strongest image models with the same prompt: Nano Banana 2, GPT Image 2, Seedream 4.5, and FLUX.2 Pro. Both failures I expected stayed away: all four rendered real woven wool and a plaid that bends correctly with the drape, and one made wool so tactile you can almost feel it. This is the textile entry in our product-photography series, alongside the skincare, jewelry, supplements, makeup, food and beverage, footwear, candles, clothing, furniture, electronics, handbags, sunglasses, glassware, flowers, watches, perfume, packaging, pet products, toys, cookware, stationery, drinkware, soap, ceramics, art prints, earbuds, houseplants, knives, and automotive wheels tests and the broader best AI image model for product photography roundup.
Quick answer
- Best overall, and cheapest photoreal: Seedream 4.5. The most tactile, fibrous wool weave, a premium macro.
- Both failures stayed away: all four rendered real woven texture (not plastic) and a plaid that follows the folds correctly (not warped).
- The caveat: this was a simple buffalo-check plaid. Fine tartans, dense repeats, and your exact colorway are harder, so use a reference and test your specific pattern.
If you only remember one thing: for a common fabric like wool, weave and drape are solved, so choose on weave richness and use a reference photo when the exact pattern has to be yours.
The test, model by model
One brief, four models, same prompt. I judged the pattern across the folds first, then the weave, then the drape and fringe.
Seedream 4.5 made the most convincing wool by a distance. You can see the individual fibers and the fuzzy halo of a real woven surface, the exact thing models usually flatten into plastic, and the green-and-cream plaid follows the fold over the chair arm correctly, distorting with the cloth the way a real plaid does. The fringe is detailed and the light is warm. This is the same material strength it shows on flowers and pet fur, and fabric plays right to it. Best result, lowest cost of the photoreal options.
Nano Banana 2 rendered the crispest weave structure: the diagonal twill in the checks reads clearly, the plaid holds its grid as it drapes, and the twisted-wool fringe is detailed. It is slightly less fibrous than Seedream's macro but cleaner and very legible as a product, a strong, finished throw shot at a third of GPT's cost.
GPT Image 2 produced a clean buffalo-check throw where the pattern geometry holds across the drape and the weave reads as real wool. It is a touch less tactile than Seedream and Nano up close, but it gets the two hard things right, the texture is woven, not plastic, and the plaid follows the folds. A reliable result at the highest price.
FLUX.2 Pro gave a convincing throw at the lowest cost: the weave is real rather than plastic, the plaid drapes correctly, and the fringe reads. It is softer than the macro shots, with less individual-fiber detail, the usual FLUX tradeoff of a strong overall image with the finest texture dialed back. For a cozy catalog or social shot it is solid and cheap.
The comparison
| Model | Weave (vs plastic) | Plaid across the drape | Fringe + drape | Rough cost/image |
|---|---|---|---|---|
| Seedream 4.5 | Best, fibrous wool | Correct | Detailed | ~4.8 credits |
| Nano Banana 2 | Crisp twill | Correct | Detailed | ~9.3 credits |
| GPT Image 2 | Real, slightly softer | Correct | Tidy | ~26.4 credits |
| FLUX.2 Pro | Real, softest | Correct | Soft | ~3.6 credits |
Credit costs are first-hand from this test on Masonry; per-image rates move, so check current pricing.
Why textiles came out better than expected
Textiles have a reputation for two failures, and the useful finding is that for a common fabric the top models have largely beaten both.
The plastic-fabric tell is mostly gone. Older and cheaper tools flatten a weave into a smooth plastic surface because they average the texture. These models did the opposite, rendering visible threads and a fibrous woven surface, Seedream most of all. Fabric is a material, and material is exactly where this generation of models is strongest, the same reason fur and petals came out so well. Still worth a full-zoom check, but the plasticization is no longer the default.
The pattern held the drape. The other fear, a plaid going wavy across the folds, is the clothing test's concern applied to a draped object. For a simple buffalo-check plaid, all four kept the grid correct as the fabric folded. The honest caveat is that simple is the operative word: a fine tartan, a dense repeat, or an intricate print has far more geometry to keep aligned, and that is where drift still happens. Test your actual pattern, not a generic one.
And the pattern is still generic. Each model invented its own green-and-cream plaid with a different check scale. For your exact colorway and pattern, the design is the product, so generate from a reference of the real fabric rather than a text description.
How to shoot your textile line without a studio
The workflow is the roundup approach, applied to a material AI now handles well. Trust the weave and drape for common fabrics, and choose on texture richness, Seedream for the most tactile macro. Judge the weave at full zoom to confirm it is woven and not plastic, especially on finer fabrics like silk or linen where the tell is subtler. And for your exact pattern and colorway, feed a reference image so the design is yours, which also lets you swap colorways across a collection instantly.
With the Masonry CLI you can compare the weave across models, or pass your real fabric as a reference to keep the exact pattern:
masonry image "plaid wool throw draped over a chair, soft window light, visible weave and fringe, photoreal" --model seedream-4-5 masonry image "drape this exact fabric over a chair in a sunlit room, keep the pattern and weave" --ref ./real-fabric.png --model gemini-3.1-flash-image-preview
The bottom line
Textiles are a quiet win for AI. The two things fabric is supposed to break on, plastic-looking weave and a pattern that warps across the folds, both held up for a plaid wool throw on all four models, so the choice is texture richness: Seedream 4.5 for the most tactile wool at the lowest cost, with Nano Banana 2 and GPT Image 2 close behind. Judge the weave at full zoom, test finer or denser patterns specifically, and use a reference photo when the exact design has to be yours. See how the same fidelity-first logic plays out across every product type in our best AI image model for product photography roundup, or run your own fabrics from one place with the Masonry CLI.
