daniil-kzn/qwen-image-2

Qwen-Image-2.0

Qwen-Image-2.0 (released 2026-02-10) is one model for text-to-image and image editing: an 8B Qwen3-VL encoder plus a 7B diffusion decoder, native 2K output, instructions up to about 1k tokens. Its strengths are dense, correctly spelled text in Chinese and English (slides, posters, infographics, comics, calendars, calligraphy), photorealism with low "AI look", and multi-image edits. Officially supported prompt languages are Chinese and English.

Execution goes through fal-runner (journal, price, approval, download). Prompt craft for video start frames stays in storyboard-keyframes / kling-video-prompt; this skill covers Qwen-specific parameters and prompt rules. Anatomy, reference fidelity and part transfer for any model are in image-edit-fidelity (§4a). Dated test results live in references/observations.md; API details, sizes and limits in references/api.md; ready prompt templates in references/prompt-patterns.md.

0. Availability check (do it first in a new task)

fal marked all four fal-ai/qwen-image-2/* endpoints deprecated on 2026-08-18; the successor alibaba/qwen-image-3/* is active. The standard text-to-image endpoint still served a request on 2026-11-08, but deprecated endpoints may stop at any time. Before the first paid call of a task:

1. python3 ~/.claude/skills/fal-runner/fal_run.py models qwen-image-2 and schema <endpoint>.
2. Check references/observations.md for the last successful run date.
3. If a call fails with 404/410/"deprecated"/"not found", report it to the user and offer the alternatives from references/api.md (Alibaba Model Studio needs DASHSCOPE_API_KEY, not configured on this machine as of 2026-11-08; WaveSpeed; or Qwen-Image-3 on fal as a different model that needs its own test). Never switch the model silently: the user asked for Qwen 2.

1. Tier and endpoint

Task	fal endpoint	Price seen (2026-11-08)
Drafts, exploration, layout tries	fal-ai/qwen-image-2/text-to-image	0.035 USD/image
Final image, dense text, realism	fal-ai/qwen-image-2/pro/text-to-image	0.075 USD/image
Edit draft	fal-ai/qwen-image-2/edit	0.035 USD/image
Final edit, multi-image fusion	fal-ai/qwen-image-2/pro/edit	0.075 USD/image

Model Studio names: qwen-image-2.0 (accelerated) and qwen-image-2.0-pro, both serve generation and editing on one endpoint (references/api.md). Iterate composition and text on the standard tier, move to Pro once the prompt is settled, and compare the cost of the accepted image including retries, not the price of one output.

2. Parameters that change the result

• Size. fal defaults to square_hd = 1024x1024, a quarter of the native 2K. Always pass image_size: {"width": W, "height": H}. Total pixels between 512x512 and 2048x2048, sides in multiples of 16. 2K presets: 1:1 2048x2048, 16:9 2688x1536, 9:16 1536x2688, 4:3 2368x1728, 3:4 1728x2368, 3:2 2496x1664, 2:3 1664x2496. scripts/qwen_payload.py size <ratio> computes any other ratio. For edits without image_size the output follows the LAST input image.
• Prompt expansion (enable_prompt_expansion on fal, prompt_extend on Model Studio), default true: an LLM rewrites the prompt and adds details, style and camera words (the official example turned a one-line cat into a "Canon EOS R5, National Geographic" prompt). Default for this skill: write the full prompt yourself by the rules in §3/§4 and set it to false when the image carries exact text, a fixed layout, a series that must match, or a seed comparison. Keep true only for short exploratory prompts where variety is wanted.
• Negative prompt, at most 500 characters. Text-to-image default (the official Qwen list): low resolution, low quality, deformed limbs, malformed fingers, oversaturated colors, waxy skin, faces without detail, over-smoothed, AI-generated look, chaotic composition, blurry or distorted text, plus the anatomy list (§4a). Edits: the fidelity list plus the anatomy list (the payload script's edit default; --negative blank sends " "), plus the failure you actually saw (altered face, misspelled text). Exclusions go here; the positive prompt does not list objects with "no".
• Prompt length. fal edit: 800 characters (schema). fal text-to-image: no limit in the schema; Model Studio truncates silently after 1300 tokens; WaveSpeed allows 800 characters. qwen_payload.py warns on the limits. When space runs out keep, in order: exact text strings, their positions, subjects and spatial relations, one or two style cues.
• Count and seed. num_images 1-4 (fal text-to-image), 1-6 (fal edit, Model Studio). seed 0..2147483647; the same seed gives similar, not identical, output.
• Edit inputs. 1-3 images, order = "image 1/2/3" in the prompt, 384-5000 px per side on fal (384-3072 recommended by Alibaba), at most 10 MB, PNG without alpha (the script flattens alpha). Local files go in as "@file:/abs/path" for fal_run.py.
• Safety checker is on by default; switching it off requires account authorization.

3. Text-to-image prompt rules

These follow the official Qwen rewriting system prompts (prompt_utils.py, prompt_utils_2512.py in QwenLM/Qwen-Image) and the 2.0 launch examples. Writing the prompt this way does the rewriter's job up front, which is why expansion can stay off.

1. Natural descriptive prose in one block, not keyword lists or tag soup. Structured content (infographics, tables, comics) may use line breaks and "Row 1: ..." sections, but each item is still a descriptive phrase.
2. Name the overall style explicitly: realistic photography, casual smartphone snapshot, movie poster, flat infographic, Song-dynasty gongbi painting, 3D render, anime illustration.
3. Order: subject, appearance, clothing head to shoes, pose, gaze and hands, other objects, environment, light (direction, intensity, color temperature), camera and composition, texture and mood. Restrain detail in simple scenes; add detail where it settles a real choice.
4. People: ethnicity, gender and a specific age ("about 30", not "young"); face shape, eyes, skin tone and texture, makeup; hair color, length, cut; every garment and accessory; body posture, gaze direction, what each hand does. About 200 words for a portrait. For realism, official examples use "casual iPhone snapshot, typical indoor lighting, no staged lighting".
5. Every visible string in double quotes, exactly as it must appear, with its position, orientation (horizontal, vertical, right-to-left columns), font (sans-serif, serif, handwritten, a named calligraphy style), color, size and carrier (neon tube, chalk, embossed gold foil, embroidery). Implied text becomes concrete text: not "a list of names" but the names. Close a text image with "No other text appears in the image."; a picture without text gets "The image contains no recognizable text."
6. Keep proper nouns, brands, slogans, URLs and numbers verbatim.
7. Layouts: describe the grid or columns first, then each block in reading order with its exact text; tables and calendars cell by cell; comics as "a 4x6 grid with white dividers", then "Row 1, panel 1: ... Speech bubble: "..."". Repeat a recurring character's short identity description word for word in every panel (the official comic does this) to keep them consistent. Pictures inside a slide go in parentheses next to their node.
8. Language: write the description in English or Chinese; keep quoted strings in the language they must appear in. Other scripts are not officially supported. Cyrillic tested 2026-11-08: short common words often right, longer or rarer words garbled (2 of 3 Russian strings had wrong letters while English was exact). For Russian text: keep strings short, expect retries, check each word, and plan to fix wrong words in an editor or by an edit pass.
9. No quality-tag suffixes ("8K, masterpiece, ultra HD"): the 2.0 examples do not use them and they push toward the glossy AI look the negative prompt fights.

Templates for posters, slides, infographics, portraits, products, comics and calendars are in references/prompt-patterns.md.

4. Edit instruction rules

From the official edit-prompt enhancer and the 2.0 edit examples:

• One direct, specific instruction: verb, target, attributes, position. "Add a light-gray cat in the bottom-right corner, sitting and facing the camera", not "add an animal".
• Replacement: "Replace Y with X", then X's key visual features.
• Text: Replace "old text" with "new text", quotes kept, same language and capitalization; position, color and font only when they change.
• People: name what must stay (ethnicity, gender, age, face, hairstyle, expression, outfit). Expression and makeup changes subtle, never exaggerated.
• Style transfer: describe the target style by visible features (palette, line, texture, light), and put the style sentence after the content changes.
• Moving one part from one reference into another, head swap: image-edit-fidelity §3 (see §4a).
• Multi-image: say which image supplies what ("the girl from image 1 wears the black dress from image 2 and sits in the pose from image 3"), which image gives the background, and describe each person's clothing and pose. The output aspect follows the last image: put the base image last or set image_size.
• Preservation clause naming concrete things: "Do not change the real buildings, streets, vehicles or people; keep the photo realistic", "keep the product shape, logo and layout unchanged". With an attached reference always follow image-edit-fidelity §2 (see §4a).
• The model re-renders the whole frame. Compare unchanged areas side by side with the input; when a small local edit damaged the rest, paste only the edited region back onto the original with a feathered mask (ref_diff.py --paste-back, see §4a).
• Colorization template: "Restore and colorize the photo." Inpaint/outpaint templates from the Qwen-Image-Edit enhancer ("Perform inpainting on this image. The original caption is: ...", "Extend the image beyond its boundaries using outpainting. The original caption is: ...") are not yet verified on 2.0.

4a. Anatomy, reference fidelity, part transfer: image-edit-fidelity

The model-agnostic rules live in the image-edit-fidelity skill (~/.claude/skills/image-edit-fidelity/SKILL.md) and apply to every Qwen call with people or a reference image: intact anatomy for every person (its §1), change only what was asked with reference_inventory.md and a CHANGE/KEEP split (§2), transfer of exactly one part and head swap (§3), the check with ref_diff.py and paste-back repair. Load it alongside this skill.

How those rules map onto Qwen-Image-2.0:

• Anatomy and fidelity exclusions go into negative_prompt (max 500 characters): the payload script adds the anatomy list to the text-to-image default and uses fidelity + anatomy as the edit default.
• enable_prompt_expansion: false for every edit with a reference.
• The output follows the LAST image in image_urls: put the base reference (the target of a part transfer) last; the payload script sets the edit size from it by default.
• The edit prompt, KEEP clause included, fits into 800 characters (the script warns).
• Source crops for part transfer and head swap: at least 384 px per side.
• qwen_payload.py check warns when an edit has no named preservation clause, no "Do not add or remove any people or objects", a multi-image prompt without "only"/"nothing else", or people without anatomy and hand placement.
• Check: python3 ~/.claude/skills/image-edit-fidelity/scripts/ref_diff.py <reference> <output> --allow <box>.

5. Workflow

1. Ask only what changes the image: purpose and aspect ratio, the exact text strings and their language, photo or illustration, reference images. Otherwise take the defaults above.
2. Availability and schema check (§0), choose the tier (§1). With people or a reference image: load image-edit-fidelity and write reference_inventory.md with the CHANGE/KEEP split first.
3. Write the prompt into the project folder (prompt.txt) and build the payload: python3 ~/.claude/skills/qwen-image-2/scripts/qwen_payload.py build --endpoint <id> --prompt-file prompt.txt --ratio 16:9 --expansion off -o payload.json (edits: add --image a.png --image b.png). Fix every warning it prints.
4. Quote model, number of images, size and price and wait for approval (fal-runner rules).
5. fal_run.py submit <endpoint> payload.json --label "<what>" --wait.
6. Check every output at full size before showing (§6).
7. Show the result the way the user reviews media: an HTML page with labelled <img> tiles in the built-in browser pane (global instructions), naturalWidth > 0 verified, not SendUserFile. Copy finals out of ~/.cache/fal-jobs/ with descriptive names.
8. Iterate by changing one thing per round. Record new findings with a date in references/observations.md (failures stay listed).

6. Defect check

• Every quoted string character by character at full size (crop and zoom small text with PIL): spelling, digits, dates, punctuation, line breaks; no invented extra text, no duplicates.
• Layout: row and column counts, cell order, alignment, nothing cut by the frame edge.
• Anatomy of every person, background figures included (image-edit-fidelity §1): one head, two eyes, two arms, two hands with five fingers, two legs, every limb attached to its own body at the shoulder or hip, no fused bodies. Any violation fails the image.
• People and objects: counts, gaze targets, who stands where.
• Edits: ref_diff.py against the reference, then every line of reference_inventory.md (people count and identity, objects and positions, background, light, framing, text); output pixel size and aspect equal to the reference.
• Look: waxy or over-smoothed skin, oversaturation, wrong style drift from expansion.
• Returned size equals the requested size (fal falls back to 1024x1024 when image_size is omitted).
• Background signage and small print the prompt did not ask for come out as pseudo-text: keep them blurred or out of frame, or name their exact text.

Regenerate only images with real defects; report the defects you accepted.