Marigold V2 depth (Log-stage2), converted for fastforks

One safetensors file with everything the fastforks marigold-v2-depth model loads: the Marigold V2 depth checkpoint depth/Log-stage2 of huawei-bayerlab/marigold-v2 applied to Qwen/Qwen-Image-Edit-2509, converted by tools/convert_marigold_v2.py of fastforks without diffusers, bitsandbytes or peft at load time. Use it through fastforks.depth.load("marigold-v2-depth"); the file is not a diffusers checkpoint.

Precision. NF4 transformer (4 bit, bf16 block scales), bf16 LoRA factors, norms, biases and VAE, float32 folded vectors.

Conversion script. convert_marigold_v2.py in this repo (also tools/convert_marigold_v2.py in fastforks): pack and fold produce this file from the original Qwen-Image-Edit-2509 and Marigold V2 checkpoints.

Statement of changes (the originals are not modified; this file is derived from them):

  • The Qwen-Image-Edit-2509 transformer (60 blocks, 20 B parameters) quantised to NF4 (4 bit, blocks of 64, one bf16 scale per block, bitsandbytes' 16 code values) exactly as the Marigold V2 code quantises it at every load; the packed bytes and scales equal bitsandbytes' output.
  • Marigold V2's rank-128 depth LoRA kept as separate factors (lora_A / lora_B, bf16) next to each quantised linear (merging it into 4-bit weights destroys the update).
  • Everything constant at Marigold V2's fixed timestep (t = 499/1000) and fixed prompt folded into stored vectors: the per-block modulation vectors, the output norm's modulation, and the prompt's Qwen2.5-VL text states after the text input layer (338 tokens). The text encoder is therefore not needed.
  • Marigold V2's fine-tuned Qwen-Image VAE (encoder, decoder) as a single-frame 2-D network: of each causal 3-D conv only the last temporal tap (the one that sees the frame) is kept; the temporal time_conv layers (identity on one frame) are dropped. bf16.
  • The last block's text branch (never read) is dropped.

Licence. Apache License 2.0, as both originals (Marigold V2 code and weights: Huawei Technologies Co., Ltd.; Qwen-Image-Edit-2509: Alibaba Cloud / the Qwen team). See NOTICE for the attributions. The training data of the original fine-tuning is described in the Marigold V2 paper; its terms are not checked here.

Paper. Pavlovic, Wandel, Obukhov, Bartolomei, Davydov, Tosi, Poggi, Süsstrunk, Dai: Marigold V2: Revisiting Diffusion Transformers for Monocular Depth Estimation, ACM TOG (SIGGRAPH Asia 2026), arXiv:2609.08084.

Output. Affine-invariant log depth in [-1, 1] (log metres = a * depth + b), one rectified-flow step.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for fastforks/marigold-v2-depth-nf4

Finetuned
(60)
this model

Collection including fastforks/marigold-v2-depth-nf4

Paper for fastforks/marigold-v2-depth-nf4