singam96 · Released, not on our provider

ShadeNet 5 35M (2026): one photo in, albedo, depth and normals out

ShadeNet-5-35M is a 34,652,431-parameter model on Hugging Face, under Apache-2.0, that turns one photo into albedo, relative depth and surface-normal maps in a single forward pass on a frozen DINOv2 backbone, and it cannot be run on this site, which only chats with text models.

Released, not on our providerLast updated:

Use it now

Chat

This chat starts with Inkling 2 Small, not ShadeNet-5-35M.

≈ 1 credit per reply · See all models

New accounts get 5 free credits once, valid for 7 days. They work on the lower-priced models here; the priciest ones need credits you bought.

Start chatting — 5 free credits
Maker
singam96
Type
Image
Released
No date given
OpenRouter id
None

Checked

ShadeNet-5-35M: facts, with sources

  • The card gives no release date. Hugging Face’s data for the repository singam96/ShadeNet-5-35M shows it was created on November 14, 2026 (UTC) and last changed on November 15, 2026. The first uploads in the repository were of an earlier model, ShadeNet-3.2, and the card compares ShadeNet-5 with it. These are repository dates, not an announced launch.Source: huggingface.co · checked 2026-11-19
  • The card calls ShadeNet-5 single-image inverse rendering: one RGB photo goes in, and three maps come out: albedo (the base colour of surfaces), relative depth (grayscale, worked out per image) and a surface normal map. It is not a chat or text model. The card links a hosted demo, the ShadeNet-5 Space, where you can try it in a browser.Source: huggingface.co · checked 2026-11-19
  • Size and design as printed: 34,652,431 parameters in total, of which 12,595,855 are trainable and 22.1M belong to a frozen DINOv2 ViT-S/14 backbone from Meta AI. It is a feed-forward network (a ParallelUNet) that makes one forward pass and predicts a correction to a fixed prior built from the photo. The card says there is no iteration, no time embedding and no diffusion schedule. A separate small refiner of 0.91M parameters can clean up the maps.Source: huggingface.co · checked 2026-11-19
  • Input and output as the card and its config file give them: a 448x448 photo (the PyTorch model also runs at 896; the ONNX files are fixed at 448x448), values scaled to between -1 and 1, and seven output channels, also between -1 and 1: channels 0 to 2 albedo, channel 3 depth, channels 4 to 6 normal. Depth is relative, and normals point x right, y up, z toward the viewer. The card says this version has no shading channel, unlike ShadeNet-3.2.Source: huggingface.co · checked 2026-11-19
  • The card warns about its own numbers: the validation split is small (23 images) and every label is a pseudo-label made with Marigold V2 (depth is decoded from Marigold’s colour picture), so the figures are "indicative, not benchmark-grade". The depth loss ignores overall scale and shift, so raw depth error is not what the model is trained for. The comparison with the previous run changed backbone, loss, resolution and training length together, so the card says the gain cannot be put down to one change.Source: huggingface.co · checked 2026-11-19
  • Training, as printed: Flickr8k photos with Marigold V2 pseudo-labels (8,054 training scenes, 23 held out), 19 epochs at 448 px on one RTX A4000 at about 28 minutes per epoch. The refiner was trained separately by copying the six-pass average, on 8,024 Open Images V7 photos listed as CC BY 2.0 (7,724 for training and 300 held out) for 6,000 steps, about 25 minutes. The card says the training code is not published yet.Source: huggingface.co · checked 2026-11-19
  • The supported way to run it is the ONNX export with the scripts in the repository’s inference folder: install onnxruntime, numpy, pillow, torch and huggingface_hub, then run run_onnx.py on a folder of photos with --refine (single pass plus refiner, the recommended mode), --enhanced (six passes) or --merged (one file). The two quantised model files are 27.9 MB and 31.4 MB and the single merged file with the refiner is 35 MB, and the card says the refiner costs about 0.26 s on a CPU for the 448x448 ONNX version. The full-precision PyTorch checkpoint needs training code that is not published.Source: huggingface.co · checked 2026-11-19
  • Licence: the card says code and weights are Apache-2.0, and the DINOv2 backbone is Apache-2.0 as well (Hugging Face’s data for facebook/dinov2-small agrees, with 22,056,576 parameters). The training labels are model outputs from Marigold V2, not ground truth. The example and gallery pictures on the card are held-out Flickr8k photos shown for illustration; the card says the photographs belong to their photographers with no blanket licence, so check the original Flickr page before reusing one.Source: huggingface.co · checked 2026-11-19
  • The card has one limitation: input sizes must be multiples of 448 (448 is what it was trained at), so other sizes must be resized first. For photos that are not square, the card says the script scales the long side to 448 and fills the rest by mirroring the photo’s own edge, then cuts the padding off the maps, so a non-square photo uses fewer than 448 pixels on its short side.Source: huggingface.co · checked 2026-11-19
  • The card describes the hosted demo as a free CPU Space with no install, offering a single pass, a single pass plus refiner, or six passes. Hugging Face’s own data for that Space lists different hardware (zero-a10g) and a running status on the day we read it, so we do not state what it runs on or whether it stays free.Source: huggingface.co · checked 2026-11-19 · second-hand
  • The weights can be downloaded without a fee from Hugging Face under Apache-2.0. We found no hosted API price on the pages we read, and OpenRouter has no price because it does not list the model.Source: huggingface.co · checked 2026-11-19
  • ShadeNet-5-35M is not on OpenRouter’s model list (checked November 19, 2026). It turns a photo into three maps and this site is a text chat, so this site cannot run it.Source: openrouter.ai · checked 2026-11-19 · second-hand

Our table

ShadeNet-5-35M: ShadeNet-5 error and speed in its three modes

  • Results as printed (per-map L1 error, lower is better, on the 23-image held-out validation split at 448 px; the card’s own numbers, not tested by us): mean 0.1757 for a single pass, 0.1673 for a single pass plus the refiner and 0.1628 for six passes (the photo, its mirror image and rotations of plus and minus 4 degrees, averaged). Time per image on an A4000 in bf16 is 57 ms, 68 ms and 353 ms. The card says the refiner recovers 65% of the six-pass gain for about 12 ms extra.Source: huggingface.co · checked 2026-11-19

Maker’s numbers

ShadeNet-5-35M per-map error (L1) and time per image: single pass, with refiner, six passes

People who search for ShadeNet 5 35M mostly want to know how good the maps are and how heavy it is to run. By the card, the refiner is the cheap middle choice: a small gain in error for about 12 ms more. All of it is measured against pseudo-labels on a small split. You cannot run this model here, because this site is a text chat. For another small open-weight model from Hugging Face that does chat, see Inkling 2 Small, and for a very different kind of small model read CAEL 0.4.2.

Our table: figures as printed on the Hugging Face card (23 held-out images at 448 px, lower L1 is better, time on an A4000 in bf16), read on November 19, 2026; they are the maker’s own numbers and we did not run them · checked 2026-11-19
MeasureSingle passSingle pass + refinerSix passes
Albedo L10.18040.17620.1734
Depth L10.17540.17460.1653
Normal L10.1713 (unit-length normals)0.15090.1497
Mean L10.17570.16730.1628
Time per image (A4000, bf16)57 ms68 ms353 ms

Source: huggingface.co

Our table

ShadeNet-5-35M: ShadeNet-5 ONNX files: size against error

  • The ONNX files use weight-only block quantisation: big weights stored as 4-bit or 8-bit integers with a scale per block. The card says this shrinks the file and does not speed anything up (on a CPU it is slightly slower). The card’s pick, called MIX (ViT layers at 8 bits, the rest at 4 bits), is 31.4 MB and gives 0.1645 mean L1 with six passes against 0.1629 for the 139.4 MB full-precision file, which is not uploaded.Source: huggingface.co · checked 2026-11-19

Quantised files

ShadeNet-5-35M: ShadeNet-5 ONNX variants: file size and mean L1 with the six-pass recipe

The repository holds four ONNX files (two quantised models, the refiner and one merged file). The card says quantisation shrinks the files without making them faster. The first row below is the card’s reference, and of the quantised variants only the last two rows are uploaded.

Our table: figures as printed on the Hugging Face card (23 held-out images, six-pass recipe, lower L1 is better), read on November 19, 2026; the maker’s own numbers · checked 2026-11-19
FileFile sizeMean L1 (six passes)Uploaded
fp32 ONNX (reference)139.4 MB0.1629No
All int426.7 MB0.1766No
MIX2 (ViT fc2 at int8, rest int4)27.9 MB0.1683Yes
MIX (ViT fc1 and fc2 at int8, rest int4)31.4 MB0.1645Yes

Source: huggingface.co

Get told when it is available

Sign in and press the button: we store your account and this model, nothing else. No emails are sent yet.

Nearest available

Alternatives to ShadeNet-5-35M you can use now

Inkling 2 Small

Thinking Machines Lab

$0.30 / $1 (per million tokens)

  • Thinking Machines Lab announced Inkling 2 Small on November 5, 2026, and made it available on its Tinker platform.Source: thinkingmachines.ai · checked 2026-11-14
  • A mixture-of-experts model with 594B total parameters and 17B active, with image input and variable effort.Source: thinkingmachines.ai · checked 2026-11-14
  • Text and image input, text output; context window up to 1M tokens; each token is routed to 8 of 384 experts.Source: huggingface.co · checked 2026-11-14

Qwen3.8 Flash

Alibaba

$0.15 / $0.47 (per million tokens)

  • Alibaba’s price for qwen3.8-flash is $0.15 / $0.47 per million input / output tokens (Singapore, EU, Hong Kong and US regions).Source: docs.qwencloud.com · checked 2026-11-13
  • Listed on OpenRouter as qwen/qwen3.8-flash at $0.15 / $0.47 per million tokens.Source: openrouter.ai · checked 2026-11-13

FAQ

ShadeNet-5-35M FAQ

What is ShadeNet-5-35M?

It is a small image model for single-image inverse rendering: you give it one RGB photo and it returns the albedo (the surfaces’ base colour), a relative depth map and a surface normal map. It has 34,652,431 parameters, 22.1M of them a frozen DINOv2 ViT-S/14 backbone from Meta AI. The Hugging Face repository was created on November 14, 2026; the card gives no launch date. It is not a chat or text model.

How do I run ShadeNet-5-35M?

The card says the supported way is the ONNX export. Install onnxruntime, numpy, pillow, torch and huggingface_hub, then run run_onnx.py from the repository’s inference folder on a folder of photos, with --refine for the recommended mode. It runs on a CPU: the card gives about 0.26 s for the refiner alone at 448x448. The full-precision PyTorch weights need training code that the author has not published yet.

How accurate is ShadeNet-5-35M?

By the card, mean L1 error on 23 held-out images is 0.1757 for a single pass, 0.1673 with the refiner and 0.1628 with six passes (lower is better). The card itself calls the numbers indicative, not benchmark-grade, because the split is small and every label is a pseudo-label from Marigold V2. We did not run it, and the card gives no comparison with other published models.

What is the refiner, and do I need it?

The refiner is a small extra network of 0.91M parameters that cleans up the maps from the single pass. The card says it was trained to reproduce the six-pass average and recovers 65% of that average’s gain for about 12 ms more per image on an A4000, against 353 ms for six passes. It helps the normals most and the depth hardly at all. The card recommends it as the default mode.

What size of photo does ShadeNet-5-35M take?

The model was trained at 448x448 and the card lists multiples of 448 as the only supported sizes (448 and 896 run in PyTorch; the ONNX files are fixed at 448x448). The script keeps the whole photo by scaling the long side to 448 and filling the rest by mirroring the photo’s edge, then cuts the padding from the output. The depth it returns is relative, not metres.

Is ShadeNet-5-35M open source, and can I use the pictures on its card?

The card says code and weights are Apache-2.0, and the DINOv2 backbone is Apache-2.0 too. The training labels are Marigold V2 outputs, not ground truth. The example photos on the card come from the Flickr8k set, whose photographs belong to their photographers with no blanket licence, so check the original page before reusing any of them.

Can I try ShadeNet-5-35M on this site?

No. It is not on OpenRouter’s model list and it works on images, while this site is a text chat. The card links a hosted demo on Hugging Face instead. For a text chat, on this site a typical reply to Inkling 2 Small costs about 1 credit (a credit is about $0.01 in the $5 pack), billed by its real token cost, at least 1 credit. For another small open model from Hugging Face, read SenseNova Loop 9B.

Is this the official ShadeNet website?

No. This page is unofficial and not affiliated with or endorsed by singam96. Facts link to the pages we read.