From 332d0ed7352fa88935a5e6e5b78ae76fa6c40c99 Mon Sep 17 00:00:00 2001 From: Cyan9977 Date: Fri, 26 Sep 2025 19:13:14 +0800 Subject: [PATCH] Add files via upload --- nodes.py | 82 +++++++++++++++++++++++++++++++++++++++++++++----------- 1 file changed, 67 insertions(+), 15 deletions(-) diff --git a/nodes.py b/nodes.py index e9b75e4..9368343 100644 --- a/nodes.py +++ b/nodes.py @@ -98,7 +98,9 @@ def encode_prompts_sequential(clip, ordered_items): if not cond_list: dummy = torch.zeros(1, 1) - return dummy, None, False + # 创建正确维度的 dummy pooled tensor (SDXL 需要 2816 维) + dummy_pooled = torch.zeros(1, 2816) + return dummy, dummy_pooled, False # 对 cond 做 padding padded = [] @@ -129,6 +131,11 @@ def encode_prompts_sequential(clip, ordered_items): pooled_final = None pooled_available = False + # 如果 pooled 不可用，创建 dummy pooled tensor + if pooled_final is None: + # 创建正确维度的 dummy pooled tensor (SDXL 需要 2816 维) + pooled_final = torch.zeros(cond_final.shape[0], 2816) # [N, 2816] + return cond_final, pooled_final, pooled_available @@ -208,7 +215,9 @@ class SimplePromptQueue: if not 文件路径.strip(): dummy = torch.zeros(1, 1) - return ([[dummy, {"pooled_output": ""}]],) + # 创建正确维度的 dummy pooled tensor (SDXL 需要 2816 维) + dummy_pooled = torch.zeros(1, 2816) + return ([[dummy, {"pooled_output": dummy_pooled}]],) # 从文件读取提示词 lines = _read_lines_from_source("", 文件路径, True) @@ -216,7 +225,9 @@ class SimplePromptQueue: if not lines: dummy = torch.zeros(1, 1) - return ([[dummy, {"pooled_output": ""}]],) + # 创建正确维度的 dummy pooled tensor (SDXL 需要 2816 维) + dummy_pooled = torch.zeros(1, 2816) + return ([[dummy, {"pooled_output": dummy_pooled}]],) # 模式：incremental 每次调用消费一行，使用每个标签的持久计数器 if 运行模式 == "incremental": @@ -224,13 +235,13 @@ class SimplePromptQueue: line = lines[idx] ordered = [(0, line)] cond, pooled, has_pooled = encode_prompts_sequential(clip, ordered) - meta = {"pooled_output": pooled if has_pooled and pooled is not None else ""} + meta = {"pooled_output": pooled} return ([[cond, meta]],) # 默认：所有行一起批处理 ordered = list(enumerate(lines)) cond, pooled, has_pooled = encode_prompts_sequential(clip, ordered) - meta = {"pooled_output": pooled if has_pooled and pooled is not None else ""} + meta = {"pooled_output": pooled} return ([[cond, meta]],) @classmethod @@ -303,14 +314,53 @@ class NegativePromptQueue: lines = _apply_template_to_lines(lines, 模板) if not lines: - dummy = torch.zeros(1, 1) - return ([[dummy, {"pooled_output": ""}]],) + # 使用标准的 CLIP 编码方式处理空输入 + tokens = clip.tokenize("") + cond, pooled = clip.encode_from_tokens(tokens, return_pooled=True) + if pooled is None: + pooled = torch.zeros(1, 2816) # SDXL 需要 2816 维 + return ([[cond, {"pooled_output": pooled}]],) - # 所有行一起批处理 - ordered = list(enumerate(lines)) - cond, pooled, has_pooled = encode_prompts_sequential(clip, ordered) - meta = {"pooled_output": pooled if has_pooled and pooled is not None else ""} - return ([[cond, meta]],) + # 使用标准的 CLIP 编码方式处理每一行 + cond_list = [] + pooled_list = [] + + with torch.no_grad(): + for line in lines: + tokens = clip.tokenize(line) + cond, pooled = clip.encode_from_tokens(tokens, return_pooled=True) + + # 确保 cond 是 3D 张量 [B, T, C] + if cond.dim() == 2: + cond = cond.unsqueeze(0) + + # 确保 pooled 是 2D 张量 [B, D] + if pooled is None: + pooled = torch.zeros(1, 2816) # SDXL 需要 2816 维 + elif pooled.dim() == 1: + pooled = pooled.unsqueeze(0) + + cond_list.append(cond) + pooled_list.append(pooled) + + # 将所有条件连接起来 + if cond_list: + # 对 cond 做 padding 以确保长度一致 + max_tokens = max(c.shape[1] for c in cond_list) + padded_conds = [] + for c in cond_list: + if c.shape[1] < max_tokens: + pad_len = max_tokens - c.shape[1] + c = F.pad(c, (0, 0, 0, pad_len)) + padded_conds.append(c) + cond_final = torch.cat(padded_conds, dim=0) + pooled_final = torch.cat(pooled_list, dim=0) + else: + # 空情况 + cond_final = torch.zeros(1, 1) + pooled_final = torch.zeros(1, 2816) + + return ([[cond_final, {"pooled_output": pooled_final}]],) @classmethod def IS_CHANGED(cls, clip, 多行文本, 预设模板="None", 模板="{p}", 保存预设=False, 删除预设=False): @@ -385,7 +435,9 @@ class PromptQueue: if not lines: dummy = torch.zeros(1, 1) - return ([[dummy, {"pooled_output": ""}]],) + # 创建正确维度的 dummy pooled tensor (SDXL 需要 2816 维) + dummy_pooled = torch.zeros(1, 2816) + return ([[dummy, {"pooled_output": dummy_pooled}]],) # 模式：incremental 每次调用消费一行，使用每个标签的持久计数器 if 运行模式 == "incremental": @@ -393,13 +445,13 @@ class PromptQueue: line = lines[idx] ordered = [(0, line)] cond, pooled, has_pooled = encode_prompts_sequential(clip, ordered) - meta = {"pooled_output": pooled if has_pooled and pooled is not None else ""} + meta = {"pooled_output": pooled} return ([[cond, meta]],) # 默认：所有行一起批处理 ordered = list(enumerate(lines)) cond, pooled, has_pooled = encode_prompts_sequential(clip, ordered) - meta = {"pooled_output": pooled if has_pooled and pooled is not None else ""} + meta = {"pooled_output": pooled} return ([[cond, meta]],) @classmethod