Make videos on the computer you already own. FreeVideo runs MiniMax H3 in as little as 8 GB of VRAM and 16 GB of RAM, and adapts its acceleration path to your hardware.
Release 0.4.2
Bowen Xue6a8ffe47cabb
diff --git a/freevideo_engine/release_notes.json b/freevideo_engine/release_notes.jsonindex 7922bc6..a7c70bf 100644--- a/freevideo_engine/release_notes.json+++ b/freevideo_engine/release_notes.json@@ -1,41 +1,74 @@{"schema": 1,- "product_version": "0.4.1",+ "product_version": "0.4.2","en": {- "summary": "Fixes Extend on videos made earlier, and improves progress display and several agent details.",+ "summary": "Adds optional image generation models: generate and edit pictures, make several views for character cards, and let the agent use them. Also fixes several install and launcher issues.","changes": [- "Extend works again on videos made earlier, including ones whose size was changed, and the new part now uses the previous video's picture size automatically. Clicking Extend on a result or in Creations now opens Edit with the video on the timeline and a new part already placed after it, ready to generate.",- "Autopilot can start a new conversation, and earlier ones stay in a list to return to; the agent also keeps listening after it answers, so it still responds when the conversation resumes later.",- "A percentage and the time left appear as soon as a video starts; progress keeps moving while the model loads and runs smoothly to the end, on Autopilot job cards as well.",- "In Edit, an agent can start a new timeline from just a name, and writes a short note on each clip so it is clear at a glance what the clip is.",- "A video added to the timeline from Creations now brings the description it was generated from, so it can be regenerated right away.",- "A video of 16 to 29 seconds requested through an agent is now made in one generation when the computer allows it, the same as in Create.",- "When a specific camera path, placement or order of actions keeps coming out wrong, the agent can offer to make a rough camera draft first and generate from it.",- "The question about freeing the disk space used by old model files now appears after updating.",- "At smaller sizes, more of the model stays in GPU memory, so computers with less memory read less from the drive while generating.",- "When the network is slow or a step takes a while, the install page now shows what it is doing.",- "Third-party download during setup now lists only the files this computer needs, and the list can be exported; after downloading, choose the folder and the files that pass the check are used directly.",- "When other programs use a large part of the GPU during a generation, or the GPU overheats, a short note appears under the result."+ "Adds optional image generation models. \"Generate with AI\" in the media panel makes a picture from one sentence on a blank canvas, and \"Edit this\" in the image editor changes the current picture, with up to two more pictures from your media as references.",+ "With the image generation models installed, character cards start from a front view and use it as the reference for three-quarter, side, back and close-up views; scene cards are made the same way.",+ "The agent can generate or edit pictures and add them to your media, and can first make a frame with people placed as described, then generate the video from it.",+ "Setup can install the image generation models at the same time (about 16 GB), and existing installs can download them in the background from the launcher; the download pauses while a video is generating.",+ "Online installs into a folder the system code page cannot spell, such as a Chinese folder on an English system, no longer fail at the kernel check.",+ "On Linux, when the system compiler is newer than the installed GPU toolkit supports, setup now uses a compatible compiler instead of stopping while building the GPU attention component.",+ "When a Windows security feature blocks files FreeVideo needs, the error card now says so and notes that the setting may need to be changed in Windows.",+ "After a launcher error, an address or path being typed in the settings is no longer reset to its previous value.",+ "When the agent makes a film, a shot that continues the previous one is at least 10 seconds long, 10 seconds by default; a length this computer cannot make in one run is reported before generation starts.",+ "In AI Autopilot, the conversation list updates as soon as a job starts or stops, and switching conversations quickly stays on the one selected."]},"zh": {- "summary": "修复了之前生成的视频不能“接着生成”的问题,并改进了进度显示和 Agent 的几处体验。",+ "summary": "新增可选的图像生成模型,可以生成和修改图片,为角色卡生成多个角度,Agent 也能使用;并修复了几处安装和启动器的问题。","changes": [- "之前生成的视频(包括换过尺寸的)都可以“接着生成”,不再提示不能接续;新的一段会自动用上一段的画面大小。在结果或作品上点“接着生成”,会打开剪辑视图:原视频已在时间线上,后面接好了新的一段,点“生成”继续。",- "“AI 全自动”里可以开新对话,旧对话留在列表里,可以回去接着做;Agent 答完以后会一直等下一句,隔一会儿再说也有回应。",- "点“生成”后马上显示百分比和预计剩余时间;加载模型时进度照常前进,平稳走到结束,“AI 全自动”的任务卡也一样。",- "剪辑视图里,Agent 只凭一个名字就能新建时间线,并会在每一段上写一句说明,一眼就能看出这一段是什么。",- "从“作品”放上时间线的视频会带上当时的生成描述,可以直接重新生成。",- "通过 Agent 生成 16–29 秒的视频时,电脑条件允许就一次生成,和创作面板一致。",- "指定的镜头路线、位置或动作顺序总是生成得不对时,Agent 会提议先做一个粗略的镜头草稿,再照草稿生成。",- "更新后询问是否释放旧模型文件所占硬盘空间的提示,现在会正常出现。",- "画面较小时,更多模型留在显存里,内存较小的电脑生成时读取硬盘更少。",- "安装时网络较慢或某一步较久时,安装页会显示当前正在做什么。",- "安装时选第三方下载,现在只列出这台电脑需要的文件,并可以导出下载清单;下载完成后选中那个文件夹,校验通过的文件直接使用。",- "生成期间其他程序占用了较多显卡,或显卡过热时,结果下方会出现一句简短的提示。"+ "新增可选的图像生成模型。素材面板的“AI 生成”可以在空白画布上用一句话生成一张图;图片编辑器里的“改这张”可以修改当前图片,还可以再选最多两张素材图作参考。",+ "装有图像生成模型时,角色卡先生成正面像,再以它为参考生成四分之三侧面、侧面、背面和脸部特写;场景卡也用它生成。",+ "Agent 可以生成或修改图片并放进素材,也可以先按描述的人物站位生成一张首帧,再用它生成视频。",+ "安装时可以选择同时安装图像生成模型(约 16 GB);已经安装 FreeVideo 的,可以在启动器里下载,下载在后台进行,生成视频时自动暂停。",+ "英文系统上在线安装到中文文件夹时,内核检查不再因为路径无法识别而失败。",+ "Linux 上系统编译器比 GPU 工具包支持的版本更新时,安装会换用兼容的编译器,不再停在编译 GPU 注意力加速组件这一步。",+ "Windows 的安全功能拦下运行所需的文件时,错误卡会说明原因,并提示可能需要在 Windows 中调整这项设置。",+ "启动器出错后,设置里正在输入的地址或路径不再被换回原值。",+ "Agent 拍片时,接上一镜的镜头每段至少 10 秒,没写长度时默认 10 秒;这台电脑一次生成不了的长度会在开始前说明。",+ "“AI 全自动”里任务开始或停止后,对话列表立刻更新;快速切换对话时停在所选的那一段。"]},"history": [+ {+ "product_version": "0.4.1",+ "en": {+ "summary": "Fixes Extend on videos made earlier, and improves progress display and several agent details.",+ "changes": [+ "Extend works again on videos made earlier, including ones whose size was changed, and the new part now uses the previous video's picture size automatically. Clicking Extend on a result or in Creations now opens Edit with the video on the timeline and a new part already placed after it, ready to generate.",+ "Autopilot can start a new conversation, and earlier ones stay in a list to return to; the agent also keeps listening after it answers, so it still responds when the conversation resumes later.",+ "A percentage and the time left appear as soon as a video starts; progress keeps moving while the model loads and runs smoothly to the end, on Autopilot job cards as well.",+ "In Edit, an agent can start a new timeline from just a name, and writes a short note on each clip so it is clear at a glance what the clip is.",+ "A video added to the timeline from Creations now brings the description it was generated from, so it can be regenerated right away.",+ "A video of 16 to 29 seconds requested through an agent is now made in one generation when the computer allows it, the same as in Create.",+ "When a specific camera path, placement or order of actions keeps coming out wrong, the agent can offer to make a rough camera draft first and generate from it.",+ "The question about freeing the disk space used by old model files now appears after updating.",+ "At smaller sizes, more of the model stays in GPU memory, so computers with less memory read less from the drive while generating.",+ "When the network is slow or a step takes a while, the install page now shows what it is doing.",+ "Third-party download during setup now lists only the files this computer needs, and the list can be exported; after downloading, choose the folder and the files that pass the check are used directly.",+ "When other programs use a large part of the GPU during a generation, or the GPU overheats, a short note appears under the result."+ ]+ },+ "zh": {+ "summary": "修复了之前生成的视频不能“接着生成”的问题,并改进了进度显示和 Agent 的几处体验。",+ "changes": [+ "之前生成的视频(包括换过尺寸的)都可以“接着生成”,不再提示不能接续;新的一段会自动用上一段的画面大小。在结果或作品上点“接着生成”,会打开剪辑视图:原视频已在时间线上,后面接好了新的一段,点“生成”继续。",+ "“AI 全自动”里可以开新对话,旧对话留在列表里,可以回去接着做;Agent 答完以后会一直等下一句,隔一会儿再说也有回应。",+ "点“生成”后马上显示百分比和预计剩余时间;加载模型时进度照常前进,平稳走到结束,“AI 全自动”的任务卡也一样。",+ "剪辑视图里,Agent 只凭一个名字就能新建时间线,并会在每一段上写一句说明,一眼就能看出这一段是什么。",+ "从“作品”放上时间线的视频会带上当时的生成描述,可以直接重新生成。",+ "通过 Agent 生成 16–29 秒的视频时,电脑条件允许就一次生成,和创作面板一致。",+ "指定的镜头路线、位置或动作顺序总是生成得不对时,Agent 会提议先做一个粗略的镜头草稿,再照草稿生成。",+ "更新后询问是否释放旧模型文件所占硬盘空间的提示,现在会正常出现。",+ "画面较小时,更多模型留在显存里,内存较小的电脑生成时读取硬盘更少。",+ "安装时网络较慢或某一步较久时,安装页会显示当前正在做什么。",+ "安装时选第三方下载,现在只列出这台电脑需要的文件,并可以导出下载清单;下载完成后选中那个文件夹,校验通过的文件直接使用。",+ "生成期间其他程序占用了较多显卡,或显卡过热时,结果下方会出现一句简短的提示。"+ ]+ }+ },{"product_version": "0.4.0","en": {@@ -219,44 +252,6 @@"修复:部分广告拦截列表会按文件名拦下 FreeVideo 的一个脚本,导致创作面板不显示,现在页面可以正常加载,页面检查也会指出被浏览器插件拦截的文件。在节点视图里把视频放大到全屏时,不再出现“没有完整加载”的提示。较新的 ComfyUI 不再在每个视频完成后打印资源登记警告。第二遍需要把部分数据放在内存里时(例如 8 GB Windows 显卡),本地优化可以正常运行。"]}- },- {- "product_version": "0.3.20",- "en": {- "summary": "Offline installations now open FreeVideo in ComfyUI, and on Linux, GPUs with memory to spare generate faster.",- "changes": [- "Offline installations imported from the environment package now open FreeVideo in ComfyUI: the launcher installs the FreeVideo node that the package does not carry. Before, ComfyUI started without FreeVideo and the launch stopped at \"could not load FreeVideo\". A retry now also restarts the ComfyUI that an earlier attempt left running, and the message no longer suggests repairing the environment.",- "Offline installations in a longer folder path now also get their desktop and Start menu shortcuts. Windows cannot create a shortcut to a program whose path is 260 characters or longer, and the folders of an offline installation could make the launcher copy's path that long; the copy now goes to a shorter folder in that case.",- "On Linux, when the GPU has memory to spare, attention now runs 28 or all 56 heads at a time instead of 16; how many model layers stay on the GPU, and what is kept in system memory, does not change. On a 96 GB GPU, sampling a 10-second 1344×768 video took about 11% less time; with a 16–20 GB budget the gain is about 6–9%. Smaller GPUs and Windows keep their current settings. The same seed can give a slightly different video than before, as it already does between GPUs with different memory.",- "On Linux, video decoding now uses its FP16 cache on every GPU (it was on for some only). Decoding a 10-second 1344×768 video on a 96 GB GPU took 15.4 s instead of 19.2 s and 4.2 GB less memory, with identical output."- ]- },- "zh": {- "summary": "离线安装可以在 ComfyUI 中打开 FreeVideo;在 Linux 上,显存有富余的显卡生成更快。",- "changes": [- "用环境包导入的离线安装现在可以在 ComfyUI 中打开 FreeVideo:启动器会安装环境包里没有带的 FreeVideo 节点。之前 ComfyUI 启动时没有加载 FreeVideo,启动停在“无法加载 FreeVideo”。重试时也会重启上一次留下的 ComfyUI,提示也不再让你去修复运行环境。",- "离线安装放在较长的文件夹路径下时,现在也能创建桌面和开始菜单快捷方式。Windows 无法为路径达到 260 个字符的程序创建快捷方式,而离线安装的目录层级可能让启动器副本的路径超过这个长度;这种情况下,副本会改放到路径较短的位置。",- "在 Linux 上,显存有富余时,注意力改为每次计算 28 个或全部 56 个头(原来每次 16 个);留在显卡上的模型层数和放在内存里的内容都不变。在 96 GB 显卡上,生成 10 秒 1344×768 视频的采样时间缩短约 11%;显存预算在 16–20 GB 时约缩短 6–9%。显存较小的显卡和 Windows 保持现有设置。同一个种子生成的视频可能与以前略有不同,就像在显存不同的显卡上本来就会有差异一样。",- "在 Linux 上,所有显卡的视频解码都使用 FP16 缓存(以前只有部分显卡开启)。在 96 GB 显卡上解码 10 秒 1344×768 视频从 19.2 秒降到 15.4 秒,少占 4.2 GB 显存,输出完全一致。"- ]- }- },- {- "product_version": "0.3.19",- "en": {- "summary": "On Windows, GPUs with 8 or 10 GB of memory sample the first pass a little faster when a video uses reference images, a first or last frame, or reference audio.",- "changes": [- "On Windows GPUs with 8 or 10 GB of memory, the first pass of a video up to 10 seconds long now runs attention 8 heads at a time when the video uses reference images, a first or last frame, or reference audio, as text-only videos already did, instead of 4. With one reference image, a 10-second 1344×768 video took about 8% less time per first-pass step on an 8 GB budget, staying within the same video memory budget. The same seed can give a slightly different video than before, as it already does between GPUs with different memory.",- "On Windows, the model layers kept on the GPU between first-pass steps no longer shrink from step to step, so fewer layers are transferred again on each step."- ]- },- "zh": {- "summary": "在 Windows 上,8 GB 和 10 GB 显存的显卡生成带参考图、首尾帧或参考音频的视频时,第一遍采样稍快一些。",- "changes": [- "在 8 GB 和 10 GB 显存的 Windows 显卡上,10 秒以内的视频如果带参考图、首帧或尾帧、参考音频,第一遍注意力改为每次计算 8 个头(原来是 4 个),与纯文本生成一致。带一张参考图生成 10 秒 1344×768 视频时,在 8 GB 显存预算下第一遍每步缩短约 8%,显存仍在原来的预算之内。同一个种子生成的视频可能与以前略有不同,就像在显存不同的显卡上本来就会有差异一样。",- "在 Windows 上,第一遍步与步之间保留在显卡上的模型层不再逐步减少,每步需要重新传输的层更少。"- ]- }}]}