Is there a rule of thumb to find out which versions of models will work?

[alex-ie]

I have tried running e.g. Krea-2 checkpoints from HF - some work, some do not. I have tried Singularity (MiniMax H3) - tensors mismatch.

How do determine of particular finetune/ checkpoint / file format (e.g. int8_convrot, w6a8) will load without downloading large model files?

1. For now I wonder if below will run:

https://huggingface.co/Comfy-Org/MiniMax-H3/resolve/main/diffusion_models/minimax_h3_fl2va_pruned_int8_convrot.safetensors and minimax_h3_fl2va_pruned_w6a8.safetensors. If yes to both, which will be faster / better?

2. For LTX 2.5 : links mentioned in docs lead to gated files on HF. Which files can I use instead? Can I use e.g ones by unsloth?

TIA.

[antropos17]

Treat compatibility as three separate checks: the model architecture, the checkpoint's tensor layout/names, and its quantization representation. .safetensors identifies a container; it does not guarantee that every tensor encoding in it is supported by your build.

Before a large download, compare the model card and tensor metadata with the supported-model documentation for your exact stable-diffusion.cpp revision. A ComfyUI-oriented encoding such as int8_convrot or w6a8 should not be assumed to work merely because another checkpoint from the same family works.

For the two MiniMax files, I cannot confirm loading or rank speed/quality without that loader/backend check. Please include your commit/version, backend and the first tensor-mismatch error. For LTX alternatives, also verify the expected component set and original license/access conditions; a differently packaged mirror can change the layout without changing the advertised family name.

Reference: https://github.com/leejet/stable-diffusion.cpp#readme

[Aaron3312]

You can check most of this before downloading. A .safetensors file starts with a JSON header that lists every tensor and its dtype. ComfyUI-quantized files also carry a tiny <layer>.comfy_quant tensor that holds the quant format as JSON.

from huggingface_hub import parse_safetensors_file_metadata
m = parse_safetensors_file_metadata("Comfy-Org/MiniMax-H3",
        "diffusion_models/minimax_h3_fl2va_pruned_w6a8.safetensors")
# m.tensors -> names, dtypes, shapes; only the header is fetched

Your two MiniMax files. I read their headers and one layer's comfy_quant bytes:

• minimax_h3_fl2va_pruned_int8_convrot → {"format": "int8_tensorwise", "convrot": true, "convrot_groupsize": 256}. That's exactly what docs/int8_convrot.md describes, so it should load: sd.cpp runs the INT8 weights directly. Only builds with SD_USE_UPSTREAM_GGML=ON reject it.
• minimax_h3_fl2va_pruned_w6a8 → {"format": "w6a8_int8", "group_size": 32, "convrot": true, ...}, plus extra weight_s_channel / weight_s_rel tensors. w6a8 doesn't appear anywhere in the sd.cpp sources, PRs or issues, so I'd expect it not to load. I haven't actually run it, though.

So between those two, use int8_convrot. The path the docs test is leejet's GGUFs (leejet/MiniMax-H3-GGUF, used in docs/minimax_h3.md). I can't rank int8_convrot against Q8_0 for speed or quality without benchmarking.

Rule of thumb:

• If the model has its own page in docs/ (minimax_h3.md, ltx2.md, wan.md…), use the files linked there.
• For ComfyUI-quantized safetensors, check the comfy_quant format: int8_tensorwise (with convrot) is supported, and other formats like w6a8 aren't (yet).
• Finetunes can also rename or add tensors, which gives the kind of tensor mismatch you saw with Singularity.

LTX-2.5: Lightricks/LTX-2.5 is "gated: auto". Accept the terms on the model page while logged in and access is granted immediately, then download with a token (hf auth login).

If you'd rather skip that, unsloth/LTX-2.5-GGUF isn't gated and has the files the docs ask for, with identical names:

• text_encoders/gemma4-12b-with-proj-ltx-2.5-bf16.safetensors
• vae/ltx-2.5-video-vae-conv-bf16.safetensors (the conv VAE the docs require)
• vae/ltx-2.5-audio-vae-bf16.safetensors
• latent_upscale_models/ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors

For the transformer, the docs link vantagewithai/LTX-2.5-GGUF. I haven't tested unsloth's transformer GGUFs in sd.cpp, although the LTX-2.3 section of the docs uses unsloth's. And don't swap in stock Gemma 4: the docs say it isn't a substitute.

[alex-ie]

Thanks,

1. I plan to try minimax_h3_fl2va_pruned_int8_convrot.
2. I am more newbie to sd than to LLM. For LLM, when I saw GGUF finetune of some older model, llama.cpp (koboldcpp) ran it fine. Here I saw what I considered finetunes of H3 and they did not run, e.g. https://huggingface.co/Abiray/MiniMax-H3-Singularity-GGUF/, so I was a bit frustrated. Same for Krea2 checkpoints.
3. "The path the docs test is leejet's GGUFs (leejet/MiniMax-H3-GGUF, used in docs/minimax_h3.md). I can't rank int8_convrot against Q8_0....". leejet on HF have Q4 as max quality, hence my search to try higher one.