Kandinsky 6 VSR on a DGX Spark: steady on a cartoon host, and it gave a photoreal dog a new face

Rated G

Kandinsky Lab's 2-step upscaler held a 480p cartoon host's beard still where SeedVR2 made it crawl. Then we shrank a photoreal dog clip to half size and asked it to bring it back: plain Lanczos came back closer to the original, and K6 redrew the face.

Instructional (165 s): AI hosts Daisy Lane and Rafael Duarte (AI voices, ElevenLabs) put Kandinsky 6 VSR next to SeedVR2 7B and a plain resize on the same face, show what each costs on a Spark, the beard that holds still, the mouth where it slips, and the settings that ran.
Prompt and settings

Settings: Four host segments (44.5 s of speech, eleven_v4) made with LTX 2.5 image+audio-to-video from each host's on-set still, seed 42; six voice-over panels and cut-aways in Remotion using this post's own clips; 1920x1080, 30 fps

Reference images (2)
The face at native pixels, all three playing together. Left: Lanczos. Middle: Kandinsky 6 VSR. Right: SeedVR2 7B. Watch the beard on the right crawl while the middle one holds.
Prompt and settings

Settings: 540x560 crop at x 560, y 40 of each 1664x960 output, 97 frames, stacked side by side.

The dog's head at native pixels. Left: the original. Middle: Lanczos 2x from half size. Right: Kandinsky 6 VSR 2x from half size. Look at the face in the first second, while the dog stands still, then at the head once it starts to shake.
Prompt and settings

Settings: 640x528 crop at x 400, y 60 of each 1280x704 clip, 81 frames, stacked side by side. Source: the lab's Wan 2.2 I2V A14B fp8 clip (lightx2v 4-step LoRA, 4 steps, euler, cfg 1, seed 42), 1280x704, 81 frames, 16 fps, prompt "A wet golden retriever standing on a lake dock shakes the water off its coat, droplets flying in every direction, ears flapping, in bright afternoon sun. Static camera at the dog's eye level, photorealistic, slow motion feel." Shrunk to 640x352 with ffmpeg's area filter, then upscaled 2x back to 1280x704. Lanczos: ffmpeg scale=1280:704:flags=lanczos. Kandinsky-6.0-VSR-distilled2steps-5s, kandy-sr CLI v1.1.0, x2, 2 steps, seed 42, 4 tiles, decode_spatial_splits 2, torch VAE backend. Rendered Nov 15 2026: 253 s for 5.06 s, load and compile included; 81 frames in, 81 out.

Reference images (2)
The dachshund's head at native pixels. Left: the original. Middle: Lanczos 2x from 832x480. Right: SeedVR2 7B 2x from 832x480. Watch the muzzle hair and the eye rims as the head turns to the camera.
Prompt and settings

Settings: 520x480 crop at x 840, y 160 of each 1664x960 clip, 121 frames, side by side. Source: an LTX 2.5 22B distilled (int8) image-to-video take, 1920x1088, 24 fps, 121 frames, seed 2207, prompt "Side view of a small smooth-coated brown and tan miniature dachshund standing on a light wood kitchen floor in a bright sunlit kitchen, it turns its head to look straight at the camera and blinks, then wags its tail. The camera is almost still. Photorealistic, warm morning light." Cropped to the centre 1664x960, shrunk to 832x480 with ffmpeg's area filter, then upscaled 2x back to 1664x960. Lanczos: ffmpeg scale=1664:960:flags=lanczos. SeedVR2 7B: seedvr2_7b_int8_convrot + seedvr2_ema_vae_fp16, ComfyUI core nodes, Lanczos 2x pre-resize, 21-frame chunks with overlap 1, 1 step euler/simple, cfg 1.0, seed 1, color correction none. Rendered Nov 16 2026; 121 frames in, 121 out.

Reference images (1)
All four on the dachshund's head at native pixels: the original, Lanczos, SeedVR2 7B and Kandinsky 6 VSR, the last three made from the same 832x480 copy. K6 sits between the original and SeedVR2. Watch the eye as the head turns.
Prompt and settings

Settings: 520x480 crop at x 840, y 160 of each 1664x960 clip, 121 frames, 2x2 grid. Source: an LTX 2.5 22B distilled (int8) image-to-video take, 1920x1088, 24 fps, 121 frames, seed 2207, cropped to the centre 1664x960, shrunk to 832x480 with ffmpeg's area filter, then upscaled 2x back to 1664x960. Lanczos: ffmpeg scale=1664:960:flags=lanczos. SeedVR2 7B: the settings on the clip above. Kandinsky 6 VSR: Kandinsky-6.0-VSR-distilled2steps-5s-Diffusers, kandy-sr from-video, x2, 2 steps, 9 tiles, the same eight-line config as the cartoon test. Rendered Nov 17 2026 with nothing else on the GPU; 121 frames in, 121 out; 446 s wall for 5.04 s, load and compile included.

Reference images (2)

Lanczos vs Kandinsky 6 VSR, same 97 frames

  • Clip 3: The free baseline: the same clip resized 2x with ffmpeg's Lanczos. Soft, but every frame is the source.
    Prompt and settings

    Settings: ffmpeg scale=1664:960:flags=lanczos from the 832x480 source, first 97 frames to match the K6 output.

  • Clip 2: Kandinsky 6 VSR, 2x. Look at the beard: the strands hold their shape from frame to frame instead of crawling. Then the mouth just before 2 s, where the lip line comes out a little heavier than the source's.
    Prompt and settings

    Settings: Kandinsky-6.0-VSR-distilled2steps-5s (Diffusers bundle), kandy-sr CLI v1.1.0, x2, 2 steps (fixed by the distillation), seed 42, 9 tiles, decode_spatial_splits 2, torch VAE backend. 4 s (100 frames, 25 fps) of a cartoon host made with InfiniteTalk, 832x480, upscaled 2x to 1664x960. The same clip as our SeedVR2 review. Rendered Nov 12 2026; 97 frames out.

    Reference images (3)

Kandinsky 6 VSR vs SeedVR2 7B, same 97 frames

  • Clip 4: SeedVR2 7B, 2x, from our SeedVR2 review. More hair, more texture, and you can see the beard shimmer when it plays.
    Prompt and settings

    Settings: seedvr2_7b_int8_convrot + seedvr2_ema_vae_fp16, ComfyUI 0.43 core nodes; Lanczos 2x pre-resize, 21-frame temporal chunks, overlap 1, 1 step, seed 1, color correction none. Rendered Nov 6 2026; first 97 frames shown.

Original vs Kandinsky 6 VSR from half size, the dog

  • Clip 8: The answer key: the original 1280x704 clip. We shrank this to half size and asked both upscalers to bring it back, so every frame they make can be checked against this one.
    Prompt and settings

    Settings: Source: the lab's Wan 2.2 I2V A14B fp8 clip (lightx2v 4-step LoRA, 4 steps, euler, cfg 1, seed 42), 1280x704, 81 frames, 16 fps, prompt "A wet golden retriever standing on a lake dock shakes the water off its coat, droplets flying in every direction, ears flapping, in bright afternoon sun. Static camera at the dog's eye level, photorealistic, slow motion feel." Rendered Nov 1 2026. Silent.

  • Clip 7: Kandinsky 6 VSR, 2x, from the 640x352 copy. Crisper ripples on the water than Lanczos, but watch the dog's face change while it stands, and the head blur once it shakes.
    Prompt and settings

    Settings: Kandinsky-6.0-VSR-distilled2steps-5s, kandy-sr CLI v1.1.0, x2, 2 steps, seed 42, 4 tiles, decode_spatial_splits 2, torch VAE backend. Rendered Nov 15 2026: 253 s for 5.06 s, load and compile included; 81 frames in, 81 out. Source: the lab's Wan 2.2 I2V A14B fp8 clip (lightx2v 4-step LoRA, 4 steps, euler, cfg 1, seed 42), 1280x704, 81 frames, 16 fps, prompt "A wet golden retriever standing on a lake dock shakes the water off its coat, droplets flying in every direction, ears flapping, in bright afternoon sun. Static camera at the dog's eye level, photorealistic, slow motion feel." Shrunk to 640x352 with ffmpeg's area filter, then upscaled 2x back to 1280x704.

Original vs SeedVR2 7B from 832x480, the dachshund

  • Clip 11: The answer key: the original 1664x960 crop. We shrank this to 832x480 and asked SeedVR2 7B to bring it back, so every frame it makes can be checked against this one.
    Prompt and settings

    Settings: Source: an LTX 2.5 22B distilled (int8) image-to-video take, 1920x1088, 24 fps, 121 frames, seed 2207, prompt "Side view of a small smooth-coated brown and tan miniature dachshund standing on a light wood kitchen floor in a bright sunlit kitchen, it turns its head to look straight at the camera and blinks, then wags its tail. The camera is almost still. Photorealistic, warm morning light." Cropped to the centre 1664x960. Rendered Nov 10 2026. Silent.

  • Clip 10: SeedVR2 7B, 2x, from the 832x480 copy. It looks like a sharper camera than the one that shot the original.
    Prompt and settings

    Settings: SeedVR2 7B: seedvr2_7b_int8_convrot + seedvr2_ema_vae_fp16, ComfyUI core nodes, Lanczos 2x pre-resize, 21-frame chunks with overlap 1, 1 step euler/simple, cfg 1.0, seed 1, color correction none. Rendered Nov 16 2026; 121 frames in, 121 out. Source: an LTX 2.5 22B distilled (int8) image-to-video take, 1920x1088, 24 fps, 121 frames, seed 2207, prompt "Side view of a small smooth-coated brown and tan miniature dachshund standing on a light wood kitchen floor in a bright sunlit kitchen, it turns its head to look straight at the camera and blinks, then wags its tail. The camera is almost still. Photorealistic, warm morning light." Cropped to the centre 1664x960, shrunk to 832x480 with ffmpeg's area filter, then upscaled 2x back to 1664x960.

Original vs Kandinsky 6 VSR from 832x480, the dachshund

  • Clip 13: Kandinsky 6 VSR, 2x, from the 832x480 copy. A slightly crisper version of the original, same colour.
    Prompt and settings

    Settings: Kandinsky 6 VSR: Kandinsky-6.0-VSR-distilled2steps-5s-Diffusers, kandy-sr from-video, x2, 2 steps, 9 tiles, the same eight-line config as the cartoon test. Rendered Nov 17 2026 with nothing else on the GPU; 121 frames in, 121 out; 446 s wall for 5.04 s, load and compile included. Source: an LTX 2.5 22B distilled (int8) image-to-video take, 1920x1088, 24 fps, 121 frames, seed 2207, cropped to the centre 1664x960, shrunk to 832x480 with ffmpeg's area filter, then upscaled 2x back to 1664x960.

Verdict

On a cartoon-style host, keep SeedVR2 7B as your 2x upscaler and bring in Kandinsky 6 VSR only for a held face where shimmer is the problem. There it changes 11% more than a plain resize from one frame to the next, where SeedVR2 7B changes 50% more. Keep it off photoreal footage from a small source: on a 640x352 dog it changed the face while the dog stood still and smeared the head when it shook, and a free Lanczos resize came out closer to the original on every one of the 81 frames. Before you trust any AI upscale on photoreal footage, put it next to a Lanczos resize. That goes for SeedVR2 7B too: on a photoreal 480p dachshund it handed back a sharper, slightly warmer dog than the original, with about 11 times the original's fine-edge energy on the head. Good-looking, and drawn by the model. K6 on that same 480p dachshund was the gentler of the two: a little sharper than the original, colour held, and fur that crawls far less than SeedVR2's, though Lanczos was still closer on every frame and K6 reshaped an eye mid-turn. K6 costs 89 to 174 s of compute per second of video at 832x480 in our runs and peaks around 45 GB, so it needs the box to itself.

Credits and settings

ModelKandinsky 6.0 VSR, 2-step distilled (video upscaling)
MakerKandinsky Lab
VersionKandinsky-6.0-VSR-distilled2steps-5s-Diffusers (17.5 GB bundle), kandinsky-6-sr v1.1.0 (commit b15e2aa), kandy-sr CLI
What it doesupscale
LicenseMIT
HardwareDGX Spark: NVIDIA GB10, ARM (aarch64), 128 GB unified memory. Standalone venv (not ComfyUI): torch 2.11.0+cu130, diffusers 0.42.1, torch VAE backend.
Speed174 s of compute per second of video. 2x, 832x480 to 1664x960, 2 steps, 9 tiles. Wall time for one CLI call, start to file written, including the ~2.5 min load and compile it pays on every call.
Peak memory46.1 GB. nvidia-smi per-process use of the kandy-sr process, sampled every 2 s, nothing else rendering. System free memory fell from 81.6 GB to 19.3 GB during the run.
Tested onNov 12, 2026
ReleasedSep 15, 2026
AI hosts

Daisy Lane (AI host; AI voice)

Rafael Duarte (AI host; AI voice)

Narration
Not supplied
Music source
Not supplied

What it does well

  • On the cartoon host it shimmers far less than SeedVR2: on the face it changes 11% more than a plain resize from frame to frame, against 50% for SeedVR2 7B and 78% for the 3B
  • The host's beard keeps one strand pattern across frames 20-25 instead of redrawing it (beard strip on the first clip)
  • Colour stays put: the average frame is about 2 levels brighter on every channel, with no warm or blue shift
  • Real added sharpness over Lanczos on the host's brows, eyes and beard edges: 3.4 times the fine-edge energy on the face
  • Nothing to tune: no prompt, a fixed 2 steps, and one config file of eight lines
  • From a 480p photoreal clip it adds a little bite without the crawl: 1.6 times the original's fine-edge energy on the head, and body fur that changes 13.5% more from frame to frame than the original, against 86% more for SeedVR2 7B

Where it breaks

  • Redraws a photoreal face: from a 640x352 dog clip it opened the eyes wider and changed the mouth and ears, even on frames where the dog stands still (reference strips on the dog clip). On an 832x480 dachshund it was much gentler, but still turned a blurred, half-closed eye into a clean almond shape with a hard rim while the head turned (frame 10, eye strip)
  • Smears fast motion: when the dog shakes, the head turns to mush and lands off its true position, and fur detail drops to half of what Lanczos keeps
  • Slow on GB10: 675 s for a 4 s 832x480 clip and 446 s for a 5 s one, model load and compile included each time
  • Heavy: the process peaked at 46 GB, four times SeedVR2 7B's 11.7 GB, and free memory fell by 62 GB during the run
  • Darkens the host's mouth a little: a heavier lip line than the source on frames 48-51 (mouth strip on the first clip)
  • Takes at most 5 s per call, and dropped the last 3 frames of our 100-frame clip

How we tested

An AI upscaler takes a small clip and redraws it bigger, adding detail a plain resize can't. We gave Kandinsky 6 VSR the exact clip we used for SeedVR2: 4 seconds of a cartoon host the lab made with InfiniteTalk, 832x480, the 480p that InfiniteTalk renders. We asked for 2x, so 1664x960, the same size SeedVR2 gave us. Then we lined the three results up frame for frame: plain Lanczos from ffmpeg (what you get for free), Kandinsky 6 VSR, and SeedVR2 7B.

Three measurements, all on the same 97 frames:

  • Shimmer. The average change between neighbouring frames, on the whole picture and on a native-pixel crop of the face. An upscaler that invents fresh detail every frame shows up as more change than the source has.
  • Added detail. Fine-edge energy on the face crop (the variance of a Laplacian filter), with Lanczos as 1.
  • Colour. The average difference per channel from the source.

Then we looked: frame 50 side by side, six frames of beard (20-25) and six frames of mouth (48-53). They're in the "Reference images" fold under the first clip.

What the numbers say

Fine detail on the face (Lanczos = 1)Frame-to-frame change, whole frame, vs the sourceFrame-to-frame change, face, vs LanczosColour shift (R, G, B)
Lanczos 2x1.0-2%0%none
Kandinsky 6 VSR3.4-5%+11%+1.8, +2.4, +2.5
SeedVR2 7B8.4+26%+50%+2.4, -1.3, -2.0
SeedVR2 3B14.0+43%+78%+2.2, -4.8, -6.6

That's the trade in one table. SeedVR2 adds a lot more texture, and the beard crawls because that texture is redrawn every frame. Kandinsky 6 VSR sharpens what's there and leaves it alone. On a face you'll hold on screen for a few seconds, or on any shot with hair, steady can look better than sharp. On a shot that's soft to begin with and cuts away fast, take the SeedVR2 detail.

The one place K6 changed the picture was the mouth: on frames 48-51 it drew the lip line heavier and darker than the source's. It's small at full speed, but on a lip-sync clip the mouth is the part people watch, so check it before you ship. If it shows on your clip, run that shot through SeedVR2 7B instead.

Second test: a photoreal dog, with the original as the answer key

The cartoon test showed K6 is steady. It couldn't show whether K6 is right, because there was no bigger original to check against. So we made one. We took the lab's Wan 2.2 clip of a golden retriever shaking water off on a dock (1280x704, 16 fps, 5 s), shrank it to 640x352, and asked K6 and Lanczos to bring it back to full size. Every frame they return can be laid over the frame we started from.

Lanczos won, and it wasn't close.

Lanczos 2xKandinsky 6 VSR 2x
Closeness to the original, whole frame (PSNR, higher is closer)36.2 dB26.4 dB
Fur detail on the dog's head and chest, calmest quarter of frames (original = 1)0.190.18
Fur detail on the dog's head and chest, busiest quarter of frames0.240.12

The reference strips under the dog clip show what the numbers mean. On frame 10 the dog stands still, and K6 gives it wider eyes, a different mouth and different ears: a new face built from the same pixels. On frame 70, mid-shake, it smears the head and draws it a little off where the head actually is. Lanczos is soft, but every frame is the dog that was there.

Was it the frame rate? Wan renders at 16 fps, K6 was trained at 24, and its CLI warns that a 16 fps clip is "mildly out of distribution". So we ran the same shot again at 24 fps with half the motion per frame (RIFE in-between frames, with those full-size frames as the original). The numbers got better: on the calmest frames K6 drew more fur detail than Lanczos (0.27 against 0.21), and on the shake 0.16 against Lanczos's 0.24. The face still came back redrawn (second row of the frame-40 strip). That extra detail is drawn, not recovered: those frames still sat further from the original than Lanczos's. Frame rate isn't the fix.

If you have a photoreal 480p or 720p shot, run one through K6 and a Lanczos resize side by side before you switch anything over. The cartoon came in at 480 lines and the dog at 352, and a bigger source may fare better.

The smaller clip ran faster: 4 tiles instead of 9, 253 s for 5.06 s of video, about 50 s of compute per second, load included. Memory didn't shrink with it: the process still peaked at 47.4 GB.

Third test: SeedVR2 7B on a photoreal 480p clip

The cartoon test made SeedVR2 7B our pick. The dog test sank K6 on photoreal footage. Neither told us what SeedVR2 does with photoreal footage, so we built a second answer key. We took a 1080p LTX 2.5 take of a dachshund in a sunny kitchen (it turns its head to the camera, blinks and wags), cut it to 1664x960 and shrank it to 832x480. That's the same 480 lines as the cartoon host, and the size InfiniteTalk hands you. SeedVR2 7B ran with the same settings as on the cartoon, and we laid every frame it returned over the original.

Lanczos 2xSeedVR2 7B 2x
Closeness to the original, whole frame (PSNR, higher is closer)46.1 dB30.6 dB
Closeness on the dog's head42.6 dB27.5 dB
Fine-edge energy on the head (original = 1)0.6510.9
Frame-to-frame change in the body fur, compared with the original1.5% less86% more

Here's what the reference strip under the first dachshund clip shows. SeedVR2 hands back a dog that looks shot on a better camera: separate hairs on the muzzle, hard rims round the eyes, a crisp ear edge, a coat about 2.6 levels redder and 1.6 less blue on the average frame. None of that is in the frame we started from. It's a good-looking dog. It's also the model's dog.

Lanczos is soft, and on this clip soft is close to right: the original is itself a soft AI render, so a plain resize lands almost on top of it. So pick by the job. If the shot has to cut against shots you didn't upscale, or carries a face you've locked, use Lanczos. If you want a 480p take to read like a sharper camera and you're fine with the model drawing the hair, SeedVR2 7B does that. Its new hair doesn't sit still, though: the body fur changes 86% more from frame to frame than the original's.

Memory was close to the cartoon run (11.7 GB): the 7B's ComfyUI server peaked at 12.7 GB of GPU memory, and free memory fell about 20 GB.

Fourth test: Kandinsky 6 on the same dachshund

The first dog sank K6 on photoreal footage, but that clip was tiny (640x352) and the dog was shaking. It couldn't tell us whether K6 struggles with photoreal footage or with small, fast sources. The dachshund pulls those apart: 480 lines, a slow head turn, and an answer key. We gave K6 the same 832x480 copy SeedVR2 got, with the same config as the cartoon test and nothing else on the GPU.

Lanczos 2xSeedVR2 7B 2xKandinsky 6 VSR 2x
Closeness to the original, whole frame (PSNR)46.1 dB30.6 dB33.8 dB
Closeness on the dog's head42.6 dB27.5 dB30.0 dB
Fine-edge energy on the head (original = 1)0.6510.91.6
Frame-to-frame change in the body fur, compared with the original1.5% less86% more13.5% more
Colour shift on the average frame0.3 levels or less2.6 redder, 1.6 less blueabout 0.9 brighter on every channel

At 480 lines K6 behaves. Put the four side by side in the grid clip and K6's dog is a slightly crisper copy of the original, where SeedVR2's is a different, better-lit dog. The colour stays put and the fur barely crawls. On the frames where the head moves most, K6 adds almost no detail over the original (1.06 times), so it isn't inventing texture to fill the motion.

It still draws when it's unsure. Lanczos came out closer on all 121 frames, and the eye strip shows why: at frame 10, mid-turn, the original's eye is a blurred half-lid, and K6 hands back a clean almond with a hard dark rim. When the head is still, it stays close.

So the small, shaking source looks like the bigger problem on the first dog. From 480p, if you want a little more bite with no colour change and little shimmer, K6 is the gentler of the two AI upscalers we've tried. If the face has to stay exactly as shot, Lanczos is still the answer, and it's free.

This run took 446 s for the 5.04 s clip, 88.5 s of compute per second of video, with the load and warm-up included; the 18 denoising steps took 371 s of that. That's about half the cartoon run's rate at the same size, and we haven't pinned down why. The process peaked at 44.7 GB, and free memory fell from 96 GB to 39 GB. We unloaded ComfyUI's models before the run so it had the room.

Setup and gotchas

  • Install. pip install "kandinsky-6-sr @ git+https://github.com/kandinskylab/kandinsky-6-sr" into a fresh venv, and download the 2-step distilled bundle (17.5 GB). Our script at the bottom of the page writes the config and runs it.
  • It isn't a ComfyUI node here. Kandinsky Lab ships a kandy-sr command line and a ComfyUI node; we ran the command line in its own venv so our ComfyUI servers weren't touched. torch 2.11 (cu130) and diffusers 0.42.1 installed and ran on the Spark's aarch64 with no patches.
  • It reloads on every call. Each kandy-sr run loads the 17.5 GB bundle and warms up its attention kernels (about 2.5 minutes), then compiles the video decoder before it starts on your clip. The CLI takes one file per call, so ten clips pay that ten times.
  • 5 seconds per call. The model was trained on 121 frames at 24 fps. Longer clips need cutting into 5 s pieces and joining.
  • It can drop frames off the end. Our 100-frame clip came back as 97 (the 81-frame dog came back whole). Pad the end with a few copies of the last frame if you need every one.
  • Give it the whole box. Free memory fell by 62 GB during the run. We started with 82 GB free. Before the dog runs we told our idle ComfyUI to drop its models (POST /free with unload_models: true), which took free memory from 67 to 97 GB. A ComfyUI server holding LTX 2.5 (about 42 GB in our tests) would have left it short.
  • First run writes into the model folder. It writes a small kvae.yaml next to the VAE weights. Harmless, but don't put the weights on a read-only share.

Settings

resolution_scale: 2, vae_backend: torch, seed: 42, tiles_batch_size: 1, decode_spatial_splits: 2. The distilled model always runs 2 steps (the CLI ignores --num-steps) and takes no prompt. At 832x480 it cut the picture into 9 tiles. decode_spatial_splits: 2 decodes each tile in two strips, which the README says roughly halves the decoder's peak memory; we left it on for unified memory. The full script is linked at the bottom.

Speed and memory

675 s from the command starting to the file being written, for 3.9 s of output: 174 s of compute per second of video. About 150 s of that is load and warm-up. The denoising itself took 331 s (9 tiles x 2 steps, about 18 s a step), and the rest is compiling the decoder, encoding, decoding and stitching. Kandinsky Lab's own number is 41 s for a 5 s 768x512 clip on an H100, not counting load. On a Spark, plan on minutes per clip.

SeedVR2 7B did the same clip in 330 s, load included. So K6 costs about twice the time for a steadier, softer result.

The kandy-sr process peaked at 46.1 GB (nvidia-smi, sampled every 2 s). SeedVR2 7B's server peaked at 11.7 GB on the same clip.

The script and config we ran (kandy-sr CLI, x2) ↗