🌐 AI搜索 & 代理 主页
Skip to content

Feat; Add support for Wan/Qwen TAEHV decoding #937

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Draft
stduhpf wants to merge 8 commits into
base: master
Choose a base branch
from
+368 −31
Draft

Feat; Add support for Wan/Qwen TAEHV decoding #937

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
b671126
Add support for Wan2.1 TAEHV decoding
stduhpf Dec 11, 2025
776385d
--tae instead of --taesd
stduhpf Dec 11, 2025
bd0c5e2
progress towards video support
stduhpf Dec 11, 2025
56339d4
Wan2.1 decode not crashing anymore (still broken)
stduhpf Dec 11, 2025
9d0005a
Less broken video decode + remove log spam
stduhpf Dec 11, 2025
dd75498
Taehv fixes
stduhpf Dec 11, 2025
fde734b
Adapt to lastest changes
stduhpf Dec 11, 2025
baf122d
taew2.1 encode support
stduhpf Dec 12, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion examples/cli/main.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -580,7 +580,7 @@ struct SDContextParams {
"--vae",
"path to standalone vae model",
&vae_path},
{"",
{"--tae",
"--taesd",
"path to taesd. Using Tiny AutoEncoder for fast decoding (low quality)",
&taesd_path},
Expand Down
66 changes: 43 additions & 23 deletions stable-diffusion.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -400,8 +400,8 @@ class StableDiffusionGGML {
offload_params_to_cpu,
tensor_storage_map);
diffusion_model = std::make_shared<MMDiTModel>(backend,
offload_params_to_cpu,
tensor_storage_map);
offload_params_to_cpu,
tensor_storage_map);
} else if (sd_version_is_flux(version)) {
bool is_chroma = false;
for (auto pair : tensor_storage_map) {
Expand Down Expand Up @@ -461,10 +461,10 @@ class StableDiffusionGGML {
1,
true);
diffusion_model = std::make_shared<WanModel>(backend,
offload_params_to_cpu,
tensor_storage_map,
"model.diffusion_model",
version);
offload_params_to_cpu,
tensor_storage_map,
"model.diffusion_model",
version);
if (strlen(SAFE_STR(sd_ctx_params->high_noise_diffusion_model_path)) > 0) {
high_noise_diffusion_model = std::make_shared<WanModel>(backend,
offload_params_to_cpu,
Expand Down Expand Up @@ -564,14 +564,27 @@ class StableDiffusionGGML {
}

if (sd_version_is_wan(version) || sd_version_is_qwen_image(version)) {
first_stage_model = std::make_shared<WAN::WanVAERunner>(vae_backend,
offload_params_to_cpu,
tensor_storage_map,
"first_stage_model",
vae_decode_only,
version);
first_stage_model->alloc_params_buffer();
first_stage_model->get_param_tensors(tensors, "first_stage_model");
if (!use_tiny_autoencoder) {
first_stage_model = std::make_shared<WAN::WanVAERunner>(vae_backend,
offload_params_to_cpu,
tensor_storage_map,
"first_stage_model",
vae_decode_only,
version);
first_stage_model->alloc_params_buffer();
first_stage_model->get_param_tensors(tensors, "first_stage_model");
} else {
tae_first_stage = std::make_shared<TinyVideoAutoEncoder>(vae_backend,
offload_params_to_cpu,
tensor_storage_map,
"decoder",
vae_decode_only,
version);
if (sd_ctx_params->vae_conv_direct) {
LOG_INFO("Using Conv2d direct in the tae model");
tae_first_stage->set_conv2d_direct_enabled(true);
}
}
} else if (version == VERSION_CHROMA_RADIANCE) {
first_stage_model = std::make_shared<FakeVAE>(vae_backend,
offload_params_to_cpu);
Expand All @@ -598,14 +611,13 @@ class StableDiffusionGGML {
}
first_stage_model->alloc_params_buffer();
first_stage_model->get_param_tensors(tensors, "first_stage_model");
}
if (use_tiny_autoencoder) {
tae_first_stage = std::make_shared<TinyAutoEncoder>(vae_backend,
offload_params_to_cpu,
tensor_storage_map,
"decoder.layers",
vae_decode_only,
version);
} else if (use_tiny_autoencoder) {
tae_first_stage = std::make_shared<TinyImageAutoEncoder>(vae_backend,
offload_params_to_cpu,
tensor_storage_map,
"decoder.layers",
vae_decode_only,
version);
if (sd_ctx_params->vae_conv_direct) {
LOG_INFO("Using Conv2d direct in the tae model");
tae_first_stage->set_conv2d_direct_enabled(true);
Expand Down Expand Up @@ -726,13 +738,16 @@ class StableDiffusionGGML {
unet_params_mem_size += high_noise_diffusion_model->get_params_buffer_size();
}
size_t vae_params_mem_size = 0;
LOG_DEBUG("Here");
if (!use_tiny_autoencoder || sd_ctx_params->tae_preview_only) {
vae_params_mem_size = first_stage_model->get_params_buffer_size();
}
if (use_tiny_autoencoder) {
LOG_DEBUG("Here");
if (!tae_first_stage->load_from_file(taesd_path, n_threads)) {
return false;
}
LOG_DEBUG("Here");
vae_params_mem_size = tae_first_stage->get_params_buffer_size();
}
size_t control_net_params_mem_size = 0;
Expand Down Expand Up @@ -2297,6 +2312,10 @@ class StableDiffusionGGML {
first_stage_model->free_compute_buffer();
process_vae_output_tensor(result);
} else {
if (sd_version_is_wan(version)) {
x = ggml_permute(work_ctx, x, 0, 1, 3, 2);
}

if (vae_tiling_params.enabled && !decode_video) {
// split latent in 64x64 tiles and compute in several steps
auto on_tiling = [&](ggml_tensor* in, ggml_tensor* out, bool init) {
Expand All @@ -2307,6 +2326,7 @@ class StableDiffusionGGML {
tae_first_stage->compute(n_threads, x, true, &result);
}
tae_first_stage->free_compute_buffer();

}

int64_t t1 = ggml_time_ms();
Expand Down Expand Up @@ -3817,7 +3837,7 @@ SD_API sd_image_t* generate_video(sd_ctx_t* sd_ctx, const sd_vid_gen_params_t* s
struct ggml_tensor* vid = sd_ctx->sd->decode_first_stage(work_ctx, final_latent, true);
int64_t t5 = ggml_time_ms();
LOG_INFO("decode_first_stage completed, taking %.2fs", (t5 - t4) * 1.0f / 1000);
if (sd_ctx->sd->free_params_immediately) {
if (sd_ctx->sd->free_params_immediately && !sd_ctx->sd->use_tiny_autoencoder) {
sd_ctx->sd->first_stage_model->free_params_buffer();
}

Expand Down
Loading
Loading