Prompting Qwen Image 2.1 NVFP4
Start with a concrete natural-language paragraph: medium, subject, action, relationships, composition, lighting, materials, then exact lettering. Keep each attribute next to the object it describes. The supplied workflows use 40 steps, Euler/simple, CFG 1, denoise 1, and fixed seeds.
Text to image
Describe a scene that can be depicted in one image. For a product photo, specify the object, material, label, supporting objects, surface, background and lighting. For a portrait, specify the action and framing as well as the person's appearance. The included product prompt is a complete starting point.
Set output width and height in Empty Latent Image. The encoder's resolution field controls reference-image resizing, not the independent text-to-image canvas. Changing image dimensions may change composition even with the same seed.
At CFG 1, the negative conditioning branch does not guide sampling. Put essential requirements in the positive prompt. Higher CFG is not a validated improvement for this package.
Lettering
Quote the exact text and specify its position and hierarchy. Example:
Large cream headline at the top reads exactly "TAKE THE SCENIC ROUTE". A smaller line at the bottom reads exactly "SLOW TRAVEL CLUB". Generous margins, flat navy, coral and cream inks, subtle paper grain.
Proofread the result. Clear headings and short labels worked in the release examples, but the astronaut test also shows small-label spelling errors in the BF16 baseline. Neither precision is a guarantee of perfect text.
Image editing
Upload the reference in Load Image. Connect it and the matching VAE to TextEncodeQwenImage21. Use that node's latent output for sampling, so the output matches the processed first reference's size. The included editing workflow already has those connections.
Name the changes and the elements to preserve:
Change the amber glass bottle to deep translucent cobalt-blue glass and the brass cap to brushed silver. Replace the orange with lime and rosemary with lavender. Keep the bottle shape, cream EMBER label, marble pedestal, camera angle, background and lighting.
For multiple references, refer explicitly to the first and second images. Reference preservation is semantic, not pixel-locked: small details, hand poses, clothing coverage and object placement can change.
Native transparency
Use an explicit RGBA instruction around the subject description:
This is an RGBA image with transparency. A charming hand-painted gouache sticker of a small orange fox curled around a glowing blue book, round spectacles, moss-green scarf, clean complete silhouette and generous empty margin. The image has alpha channel and the background is transparent.
Save as PNG. Check the alpha channel or composite over a checkerboard. The hidden RGB under transparent pixels can be colored; an RGB-only viewer is not an alpha test. The supplied example has alpha values spanning 0 through 255.
Comparisons
Keep the prompt, seed, dimensions, sampler, steps, backend and encoder mode fixed. The optional NVFP4 encoder patch changes conditioning arithmetic and may change the generated image. Compare stock and accelerated modes as distinct settings. The gallery contains the complete 12-prompt, two-seed comparison set.