DarthAffe/stable-diffusion.cpp
1
0
Fork 0
mirror of https://github.com/leejet/stable-diffusion.cpp.git synced 2025-12-12 21:38:58 +00:00
Code Issues Packages Projects Releases Wiki Activity

fix: allow resetting clip_skip to its default value (#697)

Browse Source
This commit is contained in:
Wagner Bruna 2025-06-30 12:23:21 -03:00 committed by GitHub
parent d6c87dce5c
commit b1fc16b504
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 30 additions and 27 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
clip.hpp
View File

@ -678,8 +678,8 @@ public:
bool with_final_ln = true; bool with_final_ln = true;
CLIPTextModel(CLIPVersion version = OPENAI_CLIP_VIT_L_14, CLIPTextModel(CLIPVersion version = OPENAI_CLIP_VIT_L_14,
int clip_skip_value = -1, bool with_final_ln = true,
bool with_final_ln = true) int clip_skip_value = -1)
: version(version), with_final_ln(with_final_ln) { : version(version), with_final_ln(with_final_ln) {
if (version == OPEN_CLIP_VIT_H_14) { if (version == OPEN_CLIP_VIT_H_14) {
hidden_size = 1024; hidden_size = 1024;
@ -701,7 +701,7 @@ public:
void set_clip_skip(int skip) { void set_clip_skip(int skip) {
if (skip <= 0) { if (skip <= 0) {
return; skip = -1;
} }
clip_skip = skip; clip_skip = skip;
} }
@ -871,9 +871,9 @@ struct CLIPTextModelRunner : public GGMLRunner {
std::map<std::string, enum ggml_type>& tensor_types, std::map<std::string, enum ggml_type>& tensor_types,
const std::string prefix, const std::string prefix,
CLIPVersion version = OPENAI_CLIP_VIT_L_14, CLIPVersion version = OPENAI_CLIP_VIT_L_14,
int clip_skip_value = 1, bool with_final_ln = true,
bool with_final_ln = true) int clip_skip_value = -1)
: GGMLRunner(backend), model(version, clip_skip_value, with_final_ln) { : GGMLRunner(backend), model(version, with_final_ln, clip_skip_value) {
model.init(params_ctx, tensor_types, prefix); model.init(params_ctx, tensor_types, prefix);
} }

45
conditioner.hpp
View File

@ -63,23 +63,24 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
PMVersion pv = PM_VERSION_1, PMVersion pv = PM_VERSION_1,
int clip_skip = -1) int clip_skip = -1)
: version(version), pm_version(pv), tokenizer(sd_version_is_sd2(version) ? 0 : 49407), embd_dir(embd_dir) { : version(version), pm_version(pv), tokenizer(sd_version_is_sd2(version) ? 0 : 49407), embd_dir(embd_dir) {
if (sd_version_is_sd1(version)) {
text_model = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "cond_stage_model.transformer.text_model", OPENAI_CLIP_VIT_L_14);
} else if (sd_version_is_sd2(version)) {
text_model = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "cond_stage_model.transformer.text_model", OPEN_CLIP_VIT_H_14);
} else if (sd_version_is_sdxl(version)) {
text_model = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "cond_stage_model.transformer.text_model", OPENAI_CLIP_VIT_L_14, false);
text_model2 = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "cond_stage_model.1.transformer.text_model", OPEN_CLIP_VIT_BIGG_14, false);
}
set_clip_skip(clip_skip);
}
void set_clip_skip(int clip_skip) {
if (clip_skip <= 0) { if (clip_skip <= 0) {
clip_skip = 1; clip_skip = 1;
if (sd_version_is_sd2(version) || sd_version_is_sdxl(version)) { if (sd_version_is_sd2(version) || sd_version_is_sdxl(version)) {
clip_skip = 2; clip_skip = 2;
} }
} }
if (sd_version_is_sd1(version)) {
text_model = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "cond_stage_model.transformer.text_model", OPENAI_CLIP_VIT_L_14, clip_skip);
} else if (sd_version_is_sd2(version)) {
text_model = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "cond_stage_model.transformer.text_model", OPEN_CLIP_VIT_H_14, clip_skip);
} else if (sd_version_is_sdxl(version)) {
text_model = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "cond_stage_model.transformer.text_model", OPENAI_CLIP_VIT_L_14, clip_skip, false);
text_model2 = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "cond_stage_model.1.transformer.text_model", OPEN_CLIP_VIT_BIGG_14, clip_skip, false);
}
}
void set_clip_skip(int clip_skip) {
text_model->set_clip_skip(clip_skip); text_model->set_clip_skip(clip_skip);
if (sd_version_is_sdxl(version)) { if (sd_version_is_sdxl(version)) {
text_model2->set_clip_skip(clip_skip); text_model2->set_clip_skip(clip_skip);
@ -665,15 +666,16 @@ struct SD3CLIPEmbedder : public Conditioner {
std::map<std::string, enum ggml_type>& tensor_types, std::map<std::string, enum ggml_type>& tensor_types,
int clip_skip = -1) int clip_skip = -1)
: clip_g_tokenizer(0) { : clip_g_tokenizer(0) {
if (clip_skip <= 0) { clip_l = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "text_encoders.clip_l.transformer.text_model", OPENAI_CLIP_VIT_L_14, false);
clip_skip = 2; clip_g = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "text_encoders.clip_g.transformer.text_model", OPEN_CLIP_VIT_BIGG_14, false);
}
clip_l = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "text_encoders.clip_l.transformer.text_model", OPENAI_CLIP_VIT_L_14, clip_skip, false);
clip_g = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "text_encoders.clip_g.transformer.text_model", OPEN_CLIP_VIT_BIGG_14, clip_skip, false);
t5 = std::make_shared<T5Runner>(backend, tensor_types, "text_encoders.t5xxl.transformer"); t5 = std::make_shared<T5Runner>(backend, tensor_types, "text_encoders.t5xxl.transformer");
set_clip_skip(clip_skip);
} }
void set_clip_skip(int clip_skip) { void set_clip_skip(int clip_skip) {
if (clip_skip <= 0) {
clip_skip = 2;
}
clip_l->set_clip_skip(clip_skip); clip_l->set_clip_skip(clip_skip);
clip_g->set_clip_skip(clip_skip); clip_g->set_clip_skip(clip_skip);
} }
@ -1010,14 +1012,15 @@ struct FluxCLIPEmbedder : public Conditioner {
FluxCLIPEmbedder(ggml_backend_t backend, FluxCLIPEmbedder(ggml_backend_t backend,
std::map<std::string, enum ggml_type>& tensor_types, std::map<std::string, enum ggml_type>& tensor_types,
int clip_skip = -1) { int clip_skip = -1) {
if (clip_skip <= 0) { clip_l = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "text_encoders.clip_l.transformer.text_model", OPENAI_CLIP_VIT_L_14, true);
clip_skip = 2;
}
clip_l = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "text_encoders.clip_l.transformer.text_model", OPENAI_CLIP_VIT_L_14, clip_skip, true);
t5 = std::make_shared<T5Runner>(backend, tensor_types, "text_encoders.t5xxl.transformer"); t5 = std::make_shared<T5Runner>(backend, tensor_types, "text_encoders.t5xxl.transformer");
set_clip_skip(clip_skip);
} }
void set_clip_skip(int clip_skip) { void set_clip_skip(int clip_skip) {
if (clip_skip <= 0) {
clip_skip = 2;
}
clip_l->set_clip_skip(clip_skip); clip_l->set_clip_skip(clip_skip);
} }
@ -1422,4 +1425,4 @@ struct PixArtCLIPEmbedder : public Conditioner {
} }
}; };
#endif #endif