Short answer: on five source images I compared by eye, SeedVR2 7B and 3B both returned clean, faithful 2x upscales, and the gap between them was smaller than you might expect. 7B tended to add more visible fabric texture (wool nap, tweed weave, denim grain, more loose hair strands). 3B looked smoother. Whether "more texture" is better depends on the fabric: on wool and tweed it looked like real detail, while on one pair of jeans it looked like invented crackle. Neither model repaired flaws that were already in the source.
This page compares the two variants of the [ZView upscaler](/upscaler/en) at 2x on full-length fashion portraits generated with Krea2 Turbo. Five source images were compared, from four different batches: a camel coat look, a cream boucle jacket look, a cream knit-and-lace look, a black blazer and wide-leg jeans look, and a beige trench street look. All output images I opened are 2176 x 3856 pixels; the cream knit source I opened is 1088 x 1928, so the output is exactly 2x in each direction. Both models return the same dimensions, so what differs is how detail is reconstructed, not resolution.
Start with what the sources look like
Differences between 3B and 7B only make sense if the same file goes into both. Each pair below is one saved Krea2 Turbo render sent to 3B and to 7B at the default 2x setting. The generation records for these sources list the model as Krea2 Turbo and were created on 19 to 24 July 2026. The upscale outputs carry no generation record, so I am not stating any settings, timings or VRAM for them.
Wool and knit: the clearest 7B gain
Camel coat and rib knit
Recorded source prompt:
street-style full body Pinterest outfit portrait, oversized camel wool coat layered over a cream rib knit top and tailored wide-leg trousers, leather ankle boots, structured tote bag, delicate gold jewelry, full outfit visibility, relaxed standing pose, soft studio lighting, visible wool weave and ribbed fabric texture, natural hair strands, clean beige background, neutral palette, editorial framing with strong silhouette clarity, realistic garment drape, polished fashion campaign composition, trend led fashion mood, outfit-first fashion editorial, complete styling system, visible fabric texture, accessories, shoes, bag, scene context, natural pose or walking action, 35mm editorial photography, realistic street, runway, retail, or lookbook setting
In the 3B output the coat reads as a fairly smooth, flat surface. The ribbed turtleneck, buttons, chain necklace and bag stitching are clean.
The 7B output shows a fuzzier wool nap and slightly more grain in the trousers, and a faint horizontal seam line appears near the lowest coat button. Composition and colours are unchanged. The source has a streaky discoloured band along the far right edge near the sleeve, and it survives in both outputs: the upscaler enlarged it rather than removing it.
Boucle jacket, pearls and fringe
Recorded source prompt:
head-to-toe lookbook portrait featuring a cropped ivory boucle jacket over a silk camisole, high-waist tailored trousers, slingback heels, quilted mini bag, layered pearl and gold accessories, soft defined curls with visible individual strands, luminous but natural studio skin tones, clean ivory background, refined Pinterest fashion mood, elegant posture, visible garment texture and seam structure, premium editorial composition, quiet luxury palette, high-end campaign styling, trend led fashion mood, outfit-first fashion editorial, complete styling system, visible fabric texture, accessories, shoes, bag, scene context, natural pose or walking action, 35mm editorial photography, realistic street, runway, retail, or lookbook setting
The 3B result keeps the face, curls, pearls, bag quilting and gold hardware intact. The jacket looks somewhat softer and more uniform, and the fringe at the hem and cuffs is less distinct.
The 7B result shows a more defined weave, clearer fringed edges and crisper gold trim. Curls and skin look broadly similar in both at the size I viewed; I did not crop in to count hair strands.
Cream cardigan and lace camisole: a near tie
I also opened a pair built from a cream cardigan, lace-trimmed silk camisole, brown belt and wool trousers (face cropped at the lips). Here the two outputs are very close. The lace, buttons, belt buckle and knit ribbing are clean in both. The small difference I could see is in the trousers: 7B shows a slightly more pebbled wool grain, while 3B is a little smoother. This pair is a useful reminder that the upgrade is not dramatic on every image, so I do not include its images here.
Denim: where "more texture" stopped looking right
Recorded source prompt for the black blazer and jeans look:
street-style full body fashion image of one adult model wearing a cropped black blazer over a white ribbed tank, straight blue denim, black minimal shoes, compact shoulder bag, small silver accessories, relaxed walking pose, complete outfit visible from head to toe, soft daylight on a clean city sidewalk, realistic hair detail, clear garment edges, visible tank ribbing, denim weave, blazer lapels, bag strap crossing the torso, neutral palette, editorial Pinterest-ready composition, natural texture, no logos, no text, current trend-led outfit styling with a modern Pinterest and Instagram editorial finish, outfit-first fashion editorial, complete styling system, visible fabric texture, accessories, shoes, bag, scene context, natural pose or walking action, 35mm editorial photography, realistic street, runway, retail, or lookbook setting
In the 3B output the jeans have a soft, mottled surface. The blazer, lapels, bag strap, hardware and rings are clean.
The 7B output gives the jeans a much stronger high-frequency pattern: pale, crackled, vein-like lines across both legs. To my eye it reads less like woven denim twill and more like a repeating texture, and I would call it over-interpreted. The blazer, bag and shoes are almost identical between the two. On this image the smoother 3B result is arguably the more natural one. A reviewer who only asks "which is sharper?" would pick the wrong file.
Street scene with signage, chain and flyaway hair
Recorded source prompt for the beige trench look:
street-style full body portrait of a Korean model wearing a light beige trench over a cropped white shirt, structured denim, low-profile leather shoes, compact shoulder bag, fine chain accessories, soft daylight on a clean Seongsu storefront wall, natural walking pose, visible layered outfit logic, crisp denim seams, flyaway hair strands, editorial street realism, neutral palette with blue denim contrast, Fujifilm GFX100S look, 45mm f/2.8 fashion framing, trend led fashion mood, outfit-first fashion editorial, complete styling system, visible fabric texture, accessories, shoes, bag, scene context, natural pose or walking action, 35mm editorial photography, realistic street, runway, retail, or lookbook setting
Both outputs keep the gold bag chain, the pendant necklaces, the trench buttons and belt buckle clean, with no doubled links that I could see. The 7B version shows somewhat more grain in the jeans and a few more loose hair strands on the left side; the 3B version is smoother. In both files the storefront lettering on the right is garbled, with letters that do not form real words. I cannot say from the files whether that came from the source or was changed by the upscale, because I did not compare the source crop, so treat it as a general warning: small text on walls and signs is a place to check.
What the five pairs add up to
| What I looked at | 3B | 7B | |---|---|---| | Wool coat, knit, tweed | Smooth, slightly flat | More nap and weave, looked real | | Cream knit and lace | Clean | Clean, marginally more wool grain | | Jeans (street look 1) | Soft, natural-looking | Crackle pattern that looked over-textured | | Jeans and trench (street look 2) | Smooth | More grain, a few more hair strands | | Chains, buckles, bag hardware | Clean | Clean |
Nothing in these five pairs looked broken in either model: no new distortion of faces, hands or chains that I could spot at the size I viewed. Only the two jeans pairs made me prefer the smoother file in one case.
Choosing between them
General guidance, not a measurement: 7B is the larger model and 3B the smaller, so 7B generally costs more memory and time per image. The actual numbers depend on your GPU and settings and I did not measure them, so I am not quoting any. The upscaler page shows which variants are available to you.
A practical routine based on the pairs above:
1. Check the source first. Fix edge streaks, cropping and odd areas before upscaling, because both models carried the right-edge streak through. 2. Upscale with both variants from the same saved file, using [Create](/create/en) for the source or [Prompt Lab](/promptlab/en) to refine the prompt. 3. Compare at 100 percent on textured regions: coat nap, tweed, denim, hair, chain links, any lettering. 4. Pick 7B when the extra texture looks like the real fabric, and 3B when it looks busy or when the image will be shown small.
Limits of this comparison
- Five sources and two model variants is a small sample. Close-up faces, eyelashes and pores were not covered in the files I viewed.
- Judgements are visual. I used no metric such as PSNR or LPIPS.
- Only the default 2x setting was used here. I have no data on other scale factors.
- Many more pairs exist in the original batches; I opened only those listed above and draw no conclusions from the rest.
FAQ
Is 7B always sharper-looking than 3B?
It added more texture on most pairs I viewed, but more is not always better. On one pair of jeans it produced a crackle pattern that looked less natural than 3B.
Do I have to regenerate the image to try the other model?
No. Keep the original file and send it to each model. Regenerating changes the picture and ruins the comparison.
Can an upscaler fix a bad source?
Not in these files. The streak on the coat source was still present after both models.
> Basis for this article > Models: Krea2 Turbo (source images, generated 19, 20, 22 and 23 July 2026 per generation records), SeedVR2 3B and SeedVR2 7B (2x upscales; no generation record, so no settings are claimed). I personally opened the 3B and 7B outputs for five pairs (camel coat, boucle jacket, cream cardigan, blazer and jeans, beige trench) plus the cream cardigan source. Output size 2176 x 3856 (and the cream source 1088 x 1928) was read from the files. The camel and boucle sources were opened for the earlier draft. No speed or VRAM measurements were taken.