# 模型介绍 **Seedream 4.0相比前代模型在多模态效果、速度和可用性上均实现显著突破：** * **多模态玩法拓展**：灵活支持文本、图像的组合输入，允许文生图、图生图、图像编辑、多图编辑、组图生成等创作模式，玩法创意多样。 * **风格化美感提升**：支持高度自由的艺术风格迁移，从巴洛克到赛博朋克风，风格百变，更可组合创造全新风格，美感突出。 * **逻辑理解力增强**：结合世界知识，提升了多模态输入理解，会“画”，更会先“想”，在涉及物理和时间约束、解谜填字、续写漫画等任务中，展现出推理生成能力。 * **自适应与 4K 生成**：可根据指令或参考图生成最佳比例图片，也支持用户自定义尺寸，最高分辨率从 2K 扩展至 4K 超高清。 * **推理速度跃升**：通过全新高效的架构设计，以及极致的蒸馏加速，DiT 生图的推理速度较 Seedream 3.0 提升超 10 倍。 # 智能生图 ## 快速生成 ![图片](https://portal.volccdn.com/obj/volcfe/cloud-universal-doc/upload_c425b2c7e83ba678d9dd3714cd22a942.gif =863x) ![图片](https://portal.volccdn.com/obj/volcfe/cloud-universal-doc/upload_9a80b86bd97a321544e55e3989119654.png =861x) ## LoRA敏捷生图基模链接：seedream4.0（训练版），支持基础生图（文生图、图生图）、多图编辑双模式： * 基础生成中，图生图仅首张参考图生效； * 多图编辑中，支持1\-6张参考图，且输入图不能为空； ![图片](https://portal.volccdn.com/obj/volcfe/cloud-universal-doc/upload_d3732f951ba0f05270e26defa15629ea.jpeg =860x) 更多参数调试跳转：ComfyUI调试(含LoRA) # LoRA训练 > 入口：参考截图Seed训练\-筛选基础模型Seedream4.0 ![图片](https://portal.volccdn.com/obj/volcfe/cloud-universal-doc/upload_745f10864fc54dbe1e7b45e090aa18be.png =873x) ## 训练图集与打标 |

|

|

|\ |板块

|说明

|示例

|\ | | | | |---|---|---| |

|1. 训练模式：|

|\ |训练数据集

| 1. t2i 单图模式|

|\ | | 2. i2i 成对模式||\ | |2. 图集质量要求|

|\ | | 1. 高分辨率2K以上（建议） |

|\ | | ||\ | | |> **中、英文均可** | |

|1. 打标：豆包、通用|^^|\ |数据集处理

| 1. t2i 单图打标：支持自定义、IP、风格化、人像四类pe；| |\ | || |\ | |> 推荐使用豆包，默认system prompt推荐：Generate a concise, accurate, and natural description of the visible content in the image without speculation.| |\ | || |\ | || |\ | | 2. i2i 成对打标：推荐使用Seed PE需要用户输入打标指令| |\ | | 3. 打标后，建议推理时prompt格式能与打标对齐。| |\ | | 4. | |\ | |2. 是否需要裁剪处理：| |\ | | 1. 单图无需，后端统一resize；| |\ | | 2. 成对数据需要保证每组内尺寸一致； | | ## 参数概要 |

|

|

|

|

|\ |参数模块

|参数名称（中英文）

|参数释义

|参数范围

|默认推荐

|\ | | | | | | |---|---|---|---|---| |

|

|每N步保存一个模型 |

|

|\ |保存设置

|每N步保存一个LoRA

| |100~1000

|200

|\ | || | | |\ | |

| | | |\ | |（save_model_steps）

| | | |\ | | | | | | |^^|

|单个训练任务输出模型数量 |

|

|\ | |保存模型最大数量（checkpoints_total_limit）

| |0～10

|10

|\ | | | | | | |

|

|最大训练迭代数|

|

|\ |训练参数

|✨训练步数

|> 区别于传统计算方式，seedream4会根据设定的training steps与图集数量，自适应epoch等参数；|100~50000

|3000

|\ |||> 建议调参|||\ |

|

|> 计算公式：NumTrainingStep = NumDataset(图集数量) \* Epoch / (BatchSize \* 8)| | |\ |

|（num_training_steps）

|> batchsize选择4096可以算作1/2左右，8192可以按照2/4去计算 | | |\ | | | | | | |^^|

|每个训练sample最大允许的token数目，等效于之前的batch size。这个参数太大会触发内存不足，谨慎调节 |

|

|\ | |训练token数目

| |2048~16384

|4096

|\ | || | | |\ | |

| | | |\ | |（max_token_per_batch）

| | | |\ | | | | | | |

|

|默认5e\-5 |

|

|\ |学习率与优化器设置

|DiT 学习率

| |\-

|5e\-5

|\ | || | | |\ | |

| | | |\ | | (learning_rate)

| | | |\ | | | | | | |^^|

|"adamw", "prodigy"|

|

|\ | |✨优化器

|> prodigy在某些任务收敛会更快，建议调参； |"adamw", "prodigy"

|adamw

|\ | || | | |\ | |

| | | |\ | | (optimizer)

| | | |\ | | | | | | |

|

|学习容量。对于训练图像数量建议：|

|

|\ |网络设置

|✨网络维度

|r 从 16→32 常有显著收益；r 从 64→128 收益会明显递减。|1\-512

|64

|\ | ||> 推荐8/16/32/64/128，建议调参； | | |\ | |

| | | |\ | |(Network_dim)

| | | |\ | | | | | | |^^|

|LoRA强度=alpha/dim，对基模的影响程度。|

|

|\ | |✨LORA权重缩放比例

||1\-512

|32

|\ | ||* 如果强度小于 1，表示适配权重变化更平滑、更保守；| | |\ | |

|| | |\ | ||> 常用值：等于 network_dim 或 network_dim\*1/2； | | |\ | | | | | | ## ComfyUI调试(含LoRA) **前置条件（必须符合）：** 1. Seedream4.0专用套餐 2. L20单卡 ![图片](https://portal.volccdn.com/obj/volcfe/cloud-universal-doc/upload_a80b5fcd0cde57a34ec0caa7906d2c45.png =779x) --- 插件下载地址： ![图片](https://portal.volccdn.com/obj/volcfe/cloud-universal-doc/upload_de0f418704429fac2ca3eadce9d4a077.png =796x) > 历史房间请升级comfyui版本至release_0.3.40 > 下载插件最新版本； --- **workflow下载：** > 包含v4.0s_edit未加速模型，但房间开启时间/生成时长均会拉长，谨慎使用。 |

|**示意图** |\ |名称

| |\ | | | |---|---| |

|![图片](https://portal.volccdn.com/obj/volcfe/cloud-universal-doc/upload_2d0ed57cb0dc4ed1ce905fbc8913e2ce.png =592x) |\ |Seedream4.0+LoRA

| |\ || |\ |

| |\ |文生图

| |\ | | | |

|![图片](https://portal.volccdn.com/obj/volcfe/cloud-universal-doc/upload_76253b5c305c6e97f29cf386493f7968.png) |\ |Seedream4.0+LoRA

| |\ || |\ |

| |\ |图片编辑

| |\ | | | |

|![图片](https://portal.volccdn.com/obj/volcfe/cloud-universal-doc/upload_c607bd7ae68723be15f9de8c456dc316.png =445x) |\ |Seedream4.0+LoRA

| |\ || |\ |

| |\ |图生图

| |\ | | | ### 编排逻辑 * **t2i单图** ![图片](https://portal.volccdn.com/obj/volcfe/cloud-universal-doc/upload_e8e6364239226d828edf2d3d25ce5805.png =795x) ![图片](https://portal.volccdn.com/obj/volcfe/cloud-universal-doc/upload_d1c9421773b72be001a1c2587c397b32.png =789x) * **多图编辑** ![图片](https://portal.volccdn.com/obj/volcfe/cloud-universal-doc/upload_5f054db1032064edc51efba7aed83afa.png =789x) ### ![图片](https://portal.volccdn.com/obj/volcfe/cloud-universal-doc/upload_97db0ac89c4392dec2812f7f32b2f19f.png =785x) ### 参数说明： |

|

|

|\ |参数名称

|截图

|参数及释义

|\ | | | | |---|---|---| |

|

|v4.0s_edit_6nfe（加速版）|\ |Seed4 DiT Loader

|

|v4.0s_edit_25nfe（未加速）|\ | ||**推荐：v4.0s_edit_6nfe**|\ | | | | |

|

|v4.0s_edit_6nfe（加速版）|\ |Seed4 VAE Loader

|

|

|**local_lora_name：替换为指定lora模型**|\ |Seed4 Lora Loader

|

|use_lora：始终保持为true|\ | | |**lora_weight：0～1，自主调试**|\ | | |**transformer_scale：自主调试，** 与训练lora时的Network alpha参数对齐|\ | | |![图片](https://portal.volccdn.com/obj/volcfe/cloud-universal-doc/upload_8fed04c03014858e88639c40cb6e5dad.png =795x) |\ | | |prefix：勿动，默认值base_model.model | |

|

|**推荐：jpeg**|\ |Seed4 VAE Encoder(Remote)

|

|jpeg编码精度损失会比png大一点，但是速度更快|\ | | | | |

|

|**官方提示词（推荐）：** |\ |PE System Prompt

|

|```Python|\ | ||# **Role**|\ | |

| You are a multimodal prompt engineer. Your purpose is to translate user requests into precise, structured instructions for generative visual models.|\ | |

| You will handle three distinct tasks. Based on the user's input and intent, you must first identify which of the following tasks the request belongs to: Text-to-Image Generation, Image Editing, or Visual Series Generation. After identifying the task, you will generate the specific output required for that task.|\ | |||\ | |

| # **Input**|\ | |

| - Text-to-Image Generation: A text prompt describing an image concept.|\ | | | - Image Editing: A text prompt describing the desired changes, along with one or more input images for reference and their resolution.|\ | | | - Visual Series Generation: A text prompt clearly describing the requirements for a sequence or collection of related images, optionally along with one or more input images for reference and their resolution.|\ | | ||\ | | | # **Task**|\ | | | 1. Text-to-Image Generation: Optimize the user's text prompt into a single, detailed paragraph of visual description suitable for an image generator. Preserve the core intent, specify key elements and style, and enclose any text meant to be visible in the image in quotation marks. Finally, provide a suitable image aspect ratio.|\ | | | 2. Image Editing: Analyze the user's text prompt and the input image(s), then describe each image based on the text and image. After that, generate a concise, actionable editing instruction and a detailed description of the expected final image after editing. Finally, provide a suitable image aspect ratio.|\ | | | 3. Visual Series Generation: Deconstruct the request into a series of individual image prompts when the user explicitly requests multiple separate images. Firstly, generate a special token, . After that, output a topic sentence that unifies the series, followed by a numbered list of prompts. Each prompt must detail a single image while ensuring narrative or stylistic continuity with the others.|\ | | ||\ | | | # **Output**|\ | | | 1. Text-to-Image Generation:|\ | | | Structure the output in JSON format with:|\ | | | - output: A single-paragraph optimized image prompt.|\ | | | - ratio: The recommended image aspect ratio.|\ | | | 2. Image Editing:|\ | | | Structure the output in JSON format with:|\ | | | - input1, input2, ...: Structured descriptions of each input image.|\ | | | - edit: Optimized editing instruction.|\ | | | - output: Description of the edited image.|\ | | | - ratio: The recommended image aspect ratio.|\ | | | 3. Visual Series Generation:|\ | | | Output a special token at first. After that, structure the output with a topic sentence first, followed by a numbered list of structured image descriptions.|\ | | |```|\ | | | | |

|

|采样步数，6～100，步数越高生成时间越长|\ |steps

|

||\ | ||* **选择v4.0s_edit_6nfe模型时，推荐：6**|\ | |

|* **选择v4.0s_edit_25nfe模式行，推荐：25** |\ | |

| |\ | | | | |

|^^|生成图像的高，[1024～4096]，**推荐2048及以上**|\ |height

| |> 生图尺寸会对齐到32的倍数 |\ | | | | |

|^^|生成图像的宽，[1024～4096]，**推荐2048及以上**|\ |width

| |> 生图尺寸会对齐到32的倍数 |\ | | | | |

|^^|seed值 |\ |seed

| | |\ | | | | |

|^^|默认1.15，取值范围[0, 4] |\ |max_shift

| | |\ | | | | |

|^^|t2i专用，文本对生图影响程度，该值越大则影响程度越大，取值范围[1.0, 10.0]|\ |guidance_scale

| |**推荐：3** |\ | | | | |

|^^|t2i专用,取值范围[0.0, 1.0]|\ |cfg_rescale_factor

| |**推荐：0.7** |\ | | | | > **备注：** > 当前Seed4 Sampler节点中的：single_edit_guidance_weight、single_edit_guidance_weight_image、single_min_cfg_time、single_max_cfg_time、multi_edit_guidance_weight、multi_edit_guidance_weight_image、multi_min_cfg_time、multi_max_cfg_time均**不生效**，为占位符，预期后续开放； # 🚧场景案例 > **实测案例，不同场景、不同训练集会有很大差别！需要实际训练并跑图测试。** 1. ## 风格化\-老课本效果测试 **敏捷测试结论：** t2i单图训练｜风格化场景 adamw优化器更佳，推理和打标尽可能相同语种(比如均为中文)，当前seedream4.0质量高于seedream3.0。 1. gpt打标瑕疵概率更高，会出现面部光斑、反白、画风不符的情况； 2. prodigy优化器有概率出现性转、画风不符现象，瑕疵率高； 3. 加速和非加速效果区别不大，缺少画面控制类节点，原图动作、角度保持一般； **推荐参数：** ![图片](https://portal.volccdn.com/obj/volcfe/cloud-universal-doc/upload_6b5a445a1e55b4127d917e360c27b08c.png =1200x) ![图片](https://portal.volccdn.com/obj/volcfe/cloud-universal-doc/upload_12d2a3afc88d0df570b732521d25e082.png =1204x) **打标：** ```YAML 自定义指令参考（豆包）：反推当前画面的人物、色彩、场景、服饰、风格、氛围等关键信息，直接输出中文自然语言结果，不要多余的标题或符号。优秀案例：画面中一只色彩绚丽的凤凰立于岩石上，周围树枝上栖息着喜鹊、麻雀、猫头鹰、老鹰等不同鸟类，它们或注视或鸣叫，仿佛在围观或交流。背景布满翠绿的枝叶，整体呈现出一幅百鸟聚集的生动场景。 ``` **训练图集：** 数量：36张包含单人、多人、场景、动物多画面 ![图片](https://portal.volccdn.com/obj/volcfe/cloud-universal-doc/upload_83acd1d2d6e81465a359fc0aa86d83ad.png =1219x) **效果对比：** > **seedream4.0不同打标、不同优化器完整对比：** |

|

|

|

|

|

|\ |任务名/图

|

|

|

|

|

|\ | | | | | | | |---|---|---|---|---|---| |

|

|

|

|

|

|\ |seedream3.0

|

|

|

|

|

|\ || | | | | |\ |

| | | | | |\ |中文caption

| | | | | |\ || | | | | |\ |

| | | | | |\ |lora权重：0.95

| | | | | |\ | | | | | | | |

|

|

|

|

|

|\ |seedream4.0

|

|

|

|

|

|\ || | | | | |\ |

| | | | | |\ |中文caption

| | | | | |\ || | | | | |\ |

| | | | | |\ |lora权重：0.95

| | | | | |\ | | | | | | | 2. ## 风格化\-怪奇物语效果测试 **风格测试结论：** t2i单图训练｜风格化场景推理和打标尽可能相同语种(比如均为中文)，测试发现4.0风格化整体出图效果\>3.0。(效果更稳定，画面融合更好) 1. 若lora中含有多种“概念”，更建议逐条打标。例如：“暗黑氛围+日本动画风格”。触发时也建议逐条描述风格特征， 2. 缺少控制类节点，肤色，姿态还原一般（纯依赖反推词，易出现“小黄人”） **打标示例：** * 数据100+张 * 打标方式：风格化\-豆包【自动打标】，英文标签。 * 触发提示词：（自动打标内容较多，需要长提示词稳定触发风格内容） **st80s_anime_style**（训练设置触发词）,The lines are delicate, the shadows are realistic, and the lighting and composition also use cinematic techniques. The dark, oppressive atmosphere. The strong red and crimson tones create a sense of foreboding and confrontation.An anime\-style illustration of （自动打标内容） ![图片](https://portal.volccdn.com/obj/volcfe/cloud-universal-doc/upload_c24dd0320a49d3f9ad315f65af2a80fa.png =1230x) > 不同触发词效果对比： |**任务名/图** |

|

|

|

|

|\ | |

|

|

|

|

|\ | ||||||\ | | | | | | | |---|---|---|---|---|---| |

|

|

|

|

|

|\ |seedream3.0

|

|

|

|

|

|\ |||||||\ |

| | | | | |\ |英文caption

| | | | | |\ || | | | | |\ |

| | | | | |\ |lora权重：0.95

| | | | | |\ | | | | | | | |

|

|

|

|

|

|\ |seedream4.0

|

|

|

|

|

|\ |||||||\ |

| | | | | |\ |英文caption

| | | | | |\ || | | | | |\ |

| | | | | |\ |lora权重：0.95

| | | | | |\ || | | | | |\ | | | | | | | |

|

|

|

|

|

|\ |seedream4.0

|

|

|

|

|

|\ |||||||\ |

| | | | | |\ |触发词组+极简caption

| | | | | |\ || | | | | |\ |

| | | | | |\ |lora权重：0.95

| | | | | |\ || | | | | |\ | | | | | | | # FAQ 1. ## 工作流报错：'transformer.blocks.0.attn.proj.lora_A.weight' 报错信息：error occurred when executing: Seed4ApplyLora exception_type: euler.errors.EulerError exception_message: EulerError, error_code: 201202, message: LoRA state dict key name is like 'transformer.blocks.0.attn.proj.lora_A.weight', but the prefix is set to 'base_model.model' ![图片](https://portal.volccdn.com/obj/volcfe/cloud-universal-doc/upload_7ff180a767862cc4551fad613f307dd7.jpeg =1223x)细解决方案：lora有问题，不适配seedream4，请更换lora 2. ## 如何在画布内加载LoRA？说明：底部菜单栏点击“manager”\-\-云资源库选择“模型中心”，搜索lora名称即可下载 ![图片](https://portal.volccdn.com/obj/volcfe/cloud-universal-doc/upload_3c4c22e2dbe3cc7582a8815c3a46de8d.gif =1220x)

图像生成大模型