This article interprets Li's 13-group report alongside Qwen's release documentation. Grow AI Skills did not run the model. The linked post contains the generated examples; the figures here explain reported measurements and acceptance checks.
Li's 196 outputs cover different tasks across 13 groups. They cannot be combined into one success rate: the poster, lamp replacement, panorama, and turntable each have a different pass condition. The useful comparison is between a task's requested result and the defect visible in that specific output.
Give generated text enough space in the final layout
Li reported a Chinese bakery poster with the headline, hours, and address spelled correctly. In a 2752 × 1536 nine-panel layout, the nine poses appeared, but labels inside the cells became unreliable at roughly 10 source pixels high. Text around 20 source pixels was steadier in his runs; that threshold comes from his layouts. His text-rendering examples also include a smaller miss: “Cinnamon Roll” became “Hinnamon Roll” in a menu while other lines were correct.
The second reduction happens when a 2752-pixel-wide image is shown at, say, 900 CSS pixels on a page. A 10-pixel source label then occupies only 3.3 CSS pixels; even a 20-pixel label becomes about 6.5. The model may spell that larger label correctly while the reader still cannot read it in the delivered design. For a poster or product graphic, assess spelling at source resolution and legibility at the size people will actually see. If an address, price, or legal line must be exact, reserve space for it and typeset that line in the final layout instead of baking it into a tiny generated panel.
Li's separate instruction-following example placed three red apples on the left, two green pears on the right, and an orange cat between them. That count-and-position result says little about the menu typo. For a graphic containing both scene instructions and customer-facing copy, inspect the objects and proofread the lettering as separate acceptance checks.
Check the exported alpha and the edges
Qwen's documentation describes a 64-channel RGBA autoencoder and gives an explicit prompt pattern for transparent output: “This is an RGBA image with transparency. <your description>. The image has alpha channel and the background is transparent.” The 64 channels describe the autoencoder, not a PNG with 64 color channels. Li reported that one glass image had 48.2% fully transparent pixels, 47.6% partly transparent pixels, and 4.2% opaque pixels. A control prompt that did not request transparency yielded an opaque image in his run.
Li's reported histogram is evidence of partly transparent pixels in that saved file. It cannot identify how the alpha was produced: a separate matting process could also leave partial transparency. Qwen's architecture documentation is the source for the model's native RGBA pathway. Neither source establishes whether the glass rim has a halo or whether a soft shadow survives compositing.
Save the output as PNG and verify that its decoded pixels actually use alpha; an alpha channel filled with 255 does not make an image transparent. Place the same file over dark and light backgrounds and inspect the glass rim, narrow gaps, and soft shadow at full size. Repeat the check after the site's image pipeline has processed the upload. Grow AI Skills' PNG Transparency Checker can distinguish used transparency from an opaque PNG, while halos still need a visual check.
Transparency generation and editing an existing transparent layer were different jobs in Li's report. Recoloring a sticker kept its transparent background and white die-cut border; compositing it onto a beach removed the transparent background while retaining that border. When he asked for a wizard hat on a transparent dragon, the new object collapsed into a small blue patch. For a sticker workflow, inspect recoloring, object insertion, and final compositing separately instead of treating one successful alpha histogram as proof that all three edits work.
Check removals and reflections after local edits
Li's marked wall-picture edit became a clock as requested. His marked lamp replacement had a different outcome: the plant appeared and the lamp remained. The text-only version made the same mistake. A region marker helped identify where to work, but it did not make replace mean remove the old object first. The acceptance condition has two parts: the new object is present and the old object is absent. Li's editing section shows both cases.
His sign edit exposes a separate failure. The new shop name on the wood was correct while the old name remained in a window reflection. Li also reported a whole-image mean absolute error of 5.29 and 1.32% of pixels changing by more than 32 intensity levels. The bounding box of those stronger changes covered the frame. A bounding box records the outermost changed pixels; a few distant pixels are enough to make it frame-wide. It does not show how much of the background changed or prove that VAE re-encoding was the cause. Without the input and output files, we cannot independently calculate a change rate outside the sign.
If a product shot must keep every pixel outside an approved region, treat the model's edit as a proposed patch. Keep the original file, define a binary mask around the object and any shadow that genuinely must change, then composite edited pixels inside that mask onto original pixels outside it. That last step guarantees pixel identity outside the mask; it does not guarantee a natural seam at the boundary, which still needs inspection. A reflected sign needs its own edited region. Splitting “remove the lamp” and “add the plant” into two passes is another sensible trial, but Li's post does not report a measured success rate for that workaround.
Inspect each identity in multi-reference scenes
Qwen documents support for up to ten reference images. Its own examples include a group portrait from six inputs and a room assembled from ten furnishing images; they demonstrate possible compositions without measuring fidelity. In Li's selected two-person result, he reported that round glasses and a beard remained on the man, while auburn curls and silver hoop earrings remained on the woman. He also reported a small white logo surviving on a kettle placed into a new scene. His post shows those examples, but it gives no pass rate across identities, brands, seeds, or reference counts.
For Li's two-person case, inspect both faces and the named markers rather than judging the overall scene alone. For the kettle, compare the body, handle, spout, and small mark against the source at the intended display size. A near-match to a product logo can still be wrong on a product page. Before public or client use, check the rights covering a recognizable person or mark; the model's ability to reproduce one grants no permission.
Reference count changes the cache decision on Li's 48GB setup
Li's full-bf16 pipeline ran on a reported 48GB RTX 4090 and kept about 30.2GB of weights resident. A standard RTX 4090 has 24GB, so disabling the reference cache alone cannot fit that same all-resident setup onto a stock card: the reported weight footprint already exceeds its VRAM by 6.2GB. Qwen's 7B figure describes the image transformer, alongside a Qwen3-VL encoder and RGBA autoencoder. The documentation lists CPU offloading, but that would be a different memory and speed configuration from Li's.
The cache mattered little in Li's text-only comparison: 10.9 versus 11.1 seconds. With conditioning images, he reported about 2GB of extra GPU memory per reference when cache was on and a 3.4-fold speed advantage in the range where he recommended it. Six references exhausted his 48GB configuration with cache on. His practical cutoff was cache on for up to four references, cache off from five onward; ten references with cache off took about 190 seconds and peaked at 38.4GB. Those are observed settings and outcomes for his pipeline, not a rule for other GPUs or runtimes. Li's cache comparison
Li also reported that VAE tiling held peak memory near 36.8GB as output size rose from 1024 to 2560 pixels, while generation time still grew. For a 2048-pixel output, he changed the default 256-pixel tile to 1024 pixels with a 768-pixel stride. In that tiling experiment he reported an 8.2GB memory saving for roughly 1% extra time; his post does not isolate how much of the saving came from changing tile size rather than enabling tiling. This still leaves the reported 30.2GB resident-weight floor. When comparing a deployable configuration, keep the prompt and output size fixed while recording reference count, cache state, VAE tile settings, peak VRAM, and end-to-end wall time. Li's tiling measurements
His 1024 × 1024 step measurement was roughly 0.526 seconds, so 24 and 40 denoising steps imply about 12.6 and 21.0 seconds of denoising under that condition, before encoding and other work. The 190-second ten-reference result is a different workload and cannot isolate the cache's effect. Li also reported that raising true_cfg_scale above 1 doubled compute in his run, while batches of one, two, and four images took 10.9, 22.2, and 44.2 seconds. On his setup, batching did not improve per-image throughput; another runtime needs its own timing before a queue or batch size is chosen. Li's timing data
Screen four viewpoints before generating a turntable
Li first requested numeric rotations across 24 images with wording such as rotated N degrees clockwise around her vertical axis. The subject kept facing the camera while crop and composition changed. A later eight-view run using descriptive positions, including seen from directly behind, produced the requested viewpoints with a fixed anchor image and seed. The phrase choice helped in that run; it did not guarantee every subsequent turntable. Li's viewpoint tests
The next 16-view attempt on a cartoon knight initially returned near-duplicates. Moving the viewpoint instruction within the prompt did not fix it. In Li's four-view control, the conditioned run with one seed stayed almost unchanged across front, left, back, and right; another seed produced distinct views. His eventual 16-view run turned the knight, but angular jumps and framing jitter remained. The experiment shows a seed-sensitive failure in that setup, not a general probability that any particular seed will work.
Before generating 16 positions, request front, left, back, and right from the same anchor. Check whether a distinctive feature, such as the knight's shield, actually moves around the body. If all four views face forward, try a different seed before expanding the set. Feeding each frame from the previous output risks accumulating drift: Li's chained run initially turned, then stalled while its mean absolute difference from the first frame rose to 59.45 and background brightness fell from 106 to 61.5. Keep the anchor fixed, compare crop and accessories between candidate views, and treat accepted stills as keyframes for a separate motion workflow. Li did not demonstrate a finished smooth video. His chaining and turntable results
Check panorama seams and storyboard dividers
Li's generated panorama had the broad equirectangular look but missed the left-to-right seam. His eight-pixel edge-strip comparison gave a mean absolute error of 33.71, versus 70.53 for random columns. A lower-than-random error does not make the wrap usable: put the edges adjacent or preview the image in a 360° viewer before using it as an environment map. Li's panorama example
His four-panel storyboard kept the character and narrative sequence yet omitted the requested white divider. When the boundary matters to the final layout, generate the panels for content and add the separator in the layout file; inspect the narrative and the panel geometry independently. Li's storyboard example
Resolve the use rights before adopting the released weights
The released model materials are under the Qwen Research License. It permits use of those materials for research or evaluation and requires a separate license for commercial use. That is a condition on using the released materials, not a blanket statement here about ownership of every generated image. A team planning client work or a paid product should settle the applicable model-use terms before integrating the weights. Separately, Li's benign probes included recognizable marks and characters; a successful recreation in a test is a reason to review the specific output before publication, not evidence that rights clearance is automatic.
