Gallery — every version of this photo Every version of this photo is kept. Nothing here was cooked or photographed — the images are generated on ATAILA hardware, and when the models or the prompt improve we regenerate and keep the old one so the difference is visible.
Default
Version 4 14/08/2026 Apache-2.0
Image model qwen_image_2512_bf16 [38.1 GB] Text encoder qwen_2.5_vl_7b [15.4 GB] Prompt and details Version 3 14/08/2026 Apache-2.0
Image model qwen_image_2512_bf16 [38.1 GB] Text encoder qwen_2.5_vl_7b [15.4 GB] Prompt and details Version 2 14/08/2026 Apache-2.0
Image model qwen_image_2512_fp8_e4m3fn [19.0 GB] Text encoder qwen_2.5_vl_7b [15.4 GB] Prompt and details Version 1 12/08/2026 Apache-2.0
Image model qwen_image_2512_fp8_e4m3fn [19.0 GB] Text encoder qwen_2.5_vl_7b_fp8_scaled [8.7 GB] Prompt and details Version 4 · Default Download this image
Licence: Apache-2.0 — Free to use, including commercially. No attribution required, no revenue cap.
What changed: Everything turned up: bf16 model, double the steps (40) with a higher-order sampler (res_multistep/beta), and a prompt that describes light, materials and optics instead of listing ingredients.
Image model qwen_image_2512_bf16 [38.1 GB]
Text encoder qwen_2.5_vl_7b [15.4 GB]
Generated on prod-ai-03 GPU1+GPU2 (DisTorch2 split)
Settings 40 steps · cfg 3 · seed 1304814314 The text encoder is listed because it changes what appears in the picture, not only how it looks.
Prompt Lecsó. Pepper and tomato stew. Photographed for a cookbook cover: served in a matte ceramic bowl or on a matte ceramic plate, standing on a pale oak table with visible grain. Soft natural window light from the left lays a gentle highlight across the food and casts a soft shadow to the right. Photographed on a 50mm lens at f/2.8 from a 45-degree angle — the front of the dish in sharp focus, the background falling gently out of focus. Realistic wet textures, subsurface light where the ingredients are translucent, true-to-ingredient colour. Editorial food photography, natural unsaturated colour, garnished simply and appropriately for the dish, no styling gimmicks.
Negative prompt text, watermark, logo, hands, cutlery clutter, over-garnished, impossible steam, plastic-looking sauce, glossy CGI render, 3d render, illustration, cartoon, oversaturated, HDR halo, raw when it should be cooked, blurry, low quality, deformed, extra fingers, duplicated food, duplicated bowl, second bowl, unappetising
Version 3 Download this image
Licence: Apache-2.0 — Free to use, including commercially. No attribution required, no revenue cap.
What changed: Full-precision diffusion model: 38.1 GB bf16 instead of 19.0 GB fp8, split across two RTX 3090s. Same prompt, seed and sampler as the previous version - precision is the only variable.
Image model qwen_image_2512_bf16 [38.1 GB]
Text encoder qwen_2.5_vl_7b [15.4 GB]
Generated on prod-ai-03 GPU1+GPU2 (DisTorch2 split)
Settings 20 steps · cfg 2.5 · seed 1304814314 The text encoder is listed because it changes what appears in the picture, not only how it looks.
Prompt Lecsó. Pepper and tomato stew. Photographed in a matte ceramic bowl or plate on a light oak surface, natural window light from the left, soft shadows, shallow depth of field, 45-degree angle, garnished simply and appropriately for the dish. Professional food photography, appetising, true-to-ingredient colours, realistic textures.
Negative prompt text, watermark, logo, hands, cutlery clutter, over-garnished, impossible steam, plastic-looking sauce, raw when it should be cooked, blurry, low quality, deformed, extra fingers, duplicated food, unappetising
Version 2 Download this image
Licence: Apache-2.0 — Free to use, including commercially. No attribution required, no revenue cap.
What changed: Text-encoder upgrade: same fp8 diffusion model as v1, but the prompt is now encoded by the full-precision qwen_2.5_vl_7b instead of the fp8 copy. Quantising the encoder costs prompt adherence first - this is the change that makes the image match the recipe.
Image model qwen_image_2512_fp8_e4m3fn [19.0 GB]
Text encoder qwen_2.5_vl_7b [15.4 GB]
Generated on prod-ai-03 GPU1 (bench)
Settings 20 steps · cfg 2.5 · seed 1304814314 The text encoder is listed because it changes what appears in the picture, not only how it looks.
Prompt Lecsó. Pepper and tomato stew. Photographed in a matte ceramic bowl or plate on a light oak surface, natural window light from the left, soft shadows, shallow depth of field, 45-degree angle, garnished simply and appropriately for the dish. Professional food photography, appetising, true-to-ingredient colours, realistic textures.
Negative prompt text, watermark, logo, hands, cutlery clutter, over-garnished, impossible steam, plastic-looking sauce, raw when it should be cooked, blurry, low quality, deformed, extra fingers, duplicated food, unappetising
Version 1 Download this image
Licence: Apache-2.0 — Free to use, including commercially. No attribution required, no revenue cap.
Image model qwen_image_2512_fp8_e4m3fn [19.0 GB]
Text encoder qwen_2.5_vl_7b_fp8_scaled [8.7 GB]
Generated on prod-ai-03 GPU0
Settings 20 steps · cfg 2.5 · seed 55243727 The text encoder is listed because it changes what appears in the picture, not only how it looks.
Prompt Lecsó. Pepper and tomato stew. Photographed in a matte ceramic bowl or plate on a light oak surface, natural window light from the left, soft shadows, shallow depth of field, 45-degree angle, garnished simply and appropriately for the dish. Professional food photography, appetising, true-to-ingredient colours, realistic textures.
Negative prompt text, watermark, logo, hands, cutlery clutter, over-garnished, impossible steam, plastic-looking sauce, raw when it should be cooked, blurry, low quality, deformed, extra fingers, duplicated food, unappetising
