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.
Names go back to the earlier wording: two-pass acceleration, first-pass and second-pass steps...
Bowen Xue1c31819b6faa
diff --git a/README.md b/README.mdindex 926a950..7c841d1 100644--- a/README.md+++ b/README.md@@ -39,7 +39,7 @@ It coordinates VRAM, system memory and disk, adapting weight placement, compute- **Low-memory inference**: Weight streaming, asynchronous prefetching and chunked computation keep peak memory low, enabling inference with as little as 8 GB of VRAM and 16 GB of RAM.- **Multimodal inputs**: Text prompts, first and last frames, and image, video and audio references.- **Community LoRAs**: Use MiniMax H3 LoRAs in your workflow. See [examples](docs/LoRA.md).-- **ComfyUI integration**: A dedicated creative workspace inside ComfyUI that supports two-step generation and batch generation and keeps a history of past creations. For finer control, switch to the node view to add LoRAs or customize the workflow.+- **ComfyUI integration**: A dedicated creative workspace inside ComfyUI that supports two-pass sampling and batch generation and keeps a history of past creations. For finer control, switch to the node view to add LoRAs or customize the workflow.- **One-click deployment**: The Windows, Mac and Linux launchers set up ComfyUI, the runtime environment and the models, reuse existing models, and support offline installation.## Getting Starteddiff --git a/README.zh-CN.md b/README.zh-CN.mdindex 37b1658..5d40f74 100644--- a/README.zh-CN.md+++ b/README.zh-CN.md@@ -39,7 +39,7 @@ FreeVideo 是面向消费级显卡的 MiniMax H3 本地推理引擎,基于 [Op- **低显存推理**:通过权重流式加载、异步预取与分块计算降低峰值显存,最低只需 8GB 显存和 16GB 内存。- **多模态输入**:支持文本、首帧、尾帧,以及图像、视频、音频参考输入。- **社区 LoRA**:支持在工作流中使用 MiniMax H3 社区 LoRA。[查看效果对比](docs/LoRA.zh-CN.md)。-- **ComfyUI 集成**:在 ComfyUI 中提供专门的创作面板,支持两步生成和批量生成,并可浏览历史作品;需要更精细的控制时,可切换到节点视图,接入 LoRA 或自定义工作流。+- **ComfyUI 集成**:在 ComfyUI 中提供专门的创作面板,支持二次采样和批量生成,并可浏览历史作品;需要更精细的控制时,可切换到节点视图,接入 LoRA 或自定义工作流。- **一键部署**:Windows、Mac 和 Linux 启动器自动完成 ComfyUI、运行环境与模型的部署,可复用已有模型,并支持离线安装。## 开始使用diff --git a/docs/Mac.md b/docs/Mac.mdindex 71c3980..b63cc83 100644--- a/docs/Mac.md+++ b/docs/Mac.md@@ -26,7 +26,7 @@ This approves FreeVideo only. It doesn't turn off Gatekeeper or change your othe- Before each stage, attention head groups, chunk sizes and the number of resident transformer layers are set from the unified memory available at that moment.- Macs download ConvRot int8 weights. On Apple M5 and newer, the large projections run as int8 on the GPU's Metal 4 tensor units; earlier Macs decode them to BF16 on the GPU as each layer loads.- Layers that are not kept in memory stream from disk with read-ahead.-- Text prompts, first-frame input and two-step generation are supported.+- Text prompts, first-frame input and two-pass sampling are supported.## Generation timediff --git a/docs/Mac.zh-CN.md b/docs/Mac.zh-CN.mdindex 49531d0..eb406e0 100644--- a/docs/Mac.zh-CN.md+++ b/docs/Mac.zh-CN.md@@ -26,7 +26,7 @@ Mac 预览版暂未进行 Apple 公证,因此首次打开时 macOS 会拦截- 每个阶段开始前,根据当前可用的统一内存确定注意力头分组、分块大小和常驻的 Transformer 层数。- Mac 下载 ConvRot int8 权重。M5 及更新的芯片在 GPU 的 Metal 4 张量单元上以 int8 计算大矩阵;更早的芯片在每层加载时于 GPU 上解码为 BF16 计算。- 未常驻内存的层从磁盘流式加载并提前预取。-- 支持文本、首帧输入和两步生成。+- 支持文本、首帧输入和二次采样。## 生成时间diff --git a/freevideo_engine/comfy_nodes.py b/freevideo_engine/comfy_nodes.pyindex f6143f7..8a5f8de 100644--- a/freevideo_engine/comfy_nodes.py+++ b/freevideo_engine/comfy_nodes.py@@ -42,7 +42,7 @@ class FreeVideoGenerate(io.ComfyNode):'When it is connected, the prompt and media inputs are not used. ''Conditioning from generic CLIP nodes is rejected.'),Media.Input('media', optional=True, tooltip='Unified Media panel: keyframes or ordered references.'),- io.Boolean.Input('two_pass', display_name='Two-step generation', default=True, optional=True,+ io.Boolean.Input('two_pass', display_name='Two-pass acceleration', default=True, optional=True,tooltip='Usually faster: generate at a lower resolution, then upscale and finish sampling at the target size.'),io.Int.Input('base_steps', display_name='First-pass steps', default=8, min=1, max=32, optional=True,tooltip='Default: 8. Two-pass acceleration runs only with 8 first-pass steps; any other count '@@ -51,8 +51,8 @@ class FreeVideoGenerate(io.ComfyNode):tooltip='Default: 3. Three steps use the independent refinement schedule. Changing sampling steps may reduce generation quality.'),io.Boolean.Input('force_regenerate', display_name='Force regeneration', default=False, optional=True,tooltip='Generate again even when an identical completed video is saved locally.'),- io.Boolean.Input('preview', display_name='Two-step: preview first', default=False, optional=True,- tooltip='Part of two-step generation: pause after the half-resolution first pass and save it as a preview. '+ io.Boolean.Input('preview', display_name='Two-pass: preview first pass', default=False, optional=True,+ tooltip='Part of two-pass acceleration: pause after the half-resolution first pass and save it as a preview. ''Run the second pass from the result when it looks right; it continues from the same latents.'),io.String.Input('refine_from', display_name='Second pass of preview', default='', optional=True, advanced=True,socketless=True, tooltip='Set by the Run second pass button: the preview video to continue. '@@ -93,7 +93,7 @@ class FreeVideoGenerate(io.ComfyNode):from .two_pass import validate_stepsvalidate_steps(base_steps, refine_steps, two_pass)if preview and not two_pass:- return 'Preview first needs two-step generation.'+ return 'Preview first needs two-pass acceleration.'comfy_bridge.installation()except (OSError, ValueError, KeyError) as error:from .continuation import UNSUPPORTED@@ -124,7 +124,7 @@ class FreeVideoGenerate(io.ComfyNode):notes = []if options.get('continue_from') and not two_pass:two_pass = True- notes.append('Continuing a video needs "Two-step generation", so it is on for this part.')+ notes.append('Continuing a video needs "Two-pass acceleration", so it is on for this part.')if options.get('continue_from') and preview:preview = Falsenotes.append('A continued part is made in full, without stopping at the preview.')diff --git a/freevideo_engine/continuation.py b/freevideo_engine/continuation.pyindex 22a04a4..139ab93 100644--- a/freevideo_engine/continuation.py+++ b/freevideo_engine/continuation.py@@ -35,7 +35,7 @@ SIZE_MISMATCH = 'Use the same width and height as the previous video.'# Every refusal starts with UNSUPPORTED, then says what to change.UNSUPPORTED = 'Continue from video cannot be used with these settings.'UNSUPPORTED_DEVICE = UNSUPPORTED + ' Continuing a video needs an NVIDIA GPU. To make a new video, clear "Continue from video".'-UNSUPPORTED_PASSES = UNSUPPORTED + ' Turn on "Two-step generation" and turn off the first-pass preview.'+UNSUPPORTED_PASSES = UNSUPPORTED + ' Turn on "Two-pass acceleration" and turn off the first-pass preview.'UNSUPPORTED_REFINE = UNSUPPORTED + ' Clear either "Second pass of preview" or "Continue from video".'UNSUPPORTED_MEDIA = UNSUPPORTED + ' Connect up to 3 pictures as references, and remove video, audio and pre-encoded inputs.'# A continued part keeps the cast's pictures, one per person and at most three people in a frame.diff --git a/freevideo_engine/failure_details.py b/freevideo_engine/failure_details.pyindex 2853c8f..faa5b3f 100644--- a/freevideo_engine/failure_details.py+++ b/freevideo_engine/failure_details.py@@ -452,9 +452,9 @@ def launcher_failure(value, *, zh=False, other_disk=False, disk_name='', can_rep(r'Continue from video cannot be used with these settings\. Continuing a video needs an NVIDIA', 'continuation-route',("This computer can't continue a video", '这台电脑无法接续视频'),('Continuing a video needs an NVIDIA GPU. To make a new video, clear "Continue from video".', '接续视频需要 NVIDIA 显卡。要生成新视频,请清空“Continue from video”。')),- (r'Continue from video cannot be used with these settings\. Turn on "Two-(?:pass|step)', 'continuation-passes',- ('Continuing needs two-step generation', '接续需要两步生成'),- ('Turn on "Two-step generation" and turn off the first-pass preview.', '请打开“两步生成”,并关闭首轮预览。')),+ (r'Continue from video cannot be used with these settings\. Turn on "Two-pass', 'continuation-passes',+ ('Continuing needs two-pass acceleration', '接续需要二次采样加速'),+ ('Turn on "Two-pass acceleration" and turn off the first-pass preview.', '请打开“二次采样加速”,并关闭首轮预览。')),(r'Continue from video cannot be used with these settings\. Clear either', 'continuation-refine',('A continued part is made in full', '接续会完整生成这一段'),('Clear either "Second pass of preview" or "Continue from video".', '请清空“Second pass of preview”或“Continue from video”其中一项。')),diff --git a/scripts/test_freevideo_entry.mjs b/scripts/test_freevideo_entry.mjsindex a84f6b9..0ef2313 100644--- a/scripts/test_freevideo_entry.mjs+++ b/scripts/test_freevideo_entry.mjs@@ -141,8 +141,8 @@ test('main browser entry registers both views and preserves node hooks', async tcomputeSize() { return [this.size[0], this.dom.reduce((height, row) => height + (row.options?.getMinHeight?.() || 0), 0)]; }}for (const [language, tooltip] of [- ['en', 'Changing steps may lower quality. Default: 8 first-pass and 3 refine steps. Two-step generation runs only with 8 first-pass steps. Other step counts use a single pass. Refine steps must be fewer than first-pass steps unless set to 3.'],- ['zh', '修改步数可能降低质量。默认首轮 8 步、细化 3 步;只有首轮为 8 步时才会使用两步生成,其他步数只生成一轮;细化步数不是 3 时,需少于首轮。'],+ ['en', 'Changing sampling steps may lower generation quality. Default: 8 first-pass and 3 second-pass steps. Two-pass acceleration runs only with 8 first-pass steps. Other step counts use a single pass. Other second-pass counts must be fewer than first-pass steps.'],+ ['zh', '修改采样步数可能降低生成质量。默认一采 8 步、二采 3 步;只有一采为 8 步时才会使用二次采样加速,其他步数只采样一遍;其他二采步数需小于一采步数。'],]) {await t.test(`${language} node step tooltips include the three-step exception for valid plans`, async () => {navigator.language = language;diff --git a/web/error_panel.js b/web/error_panel.jsindex 3c37633..80c16d0 100644--- a/web/error_panel.js+++ b/web/error_panel.js@@ -24,10 +24,9 @@ const PLAIN_ERRORS = [{key: 'continuation-route', message: 'Continue from video cannot be used with these settings. Continuing a video needs an NVIDIA GPU. To make a new video, clear "Continue from video".',title: ['This computer can\'t continue a video', '这台电脑无法接续视频'],summary: ['Continuing a video needs an NVIDIA GPU. To make a new video, clear "Continue from video".', '接续视频需要 NVIDIA 显卡。要生成新视频,请清空“Continue from video”。']},- {key: 'continuation-passes', message: 'Continue from video cannot be used with these settings. Turn on "Two-step generation" and turn off the first-pass preview.',- aliases: ['Continue from video cannot be used with these settings. Turn on "Two-pass acceleration" and turn off the first-pass preview.'],- title: ['Continuing needs two-step generation', '接续需要两步生成'],- summary: ['Turn on "Two-step generation" and turn off the first-pass preview.', '请打开“两步生成”,并关闭首轮预览。']},+ {key: 'continuation-passes', message: 'Continue from video cannot be used with these settings. Turn on "Two-pass acceleration" and turn off the first-pass preview.',+ title: ['Continuing needs two-pass acceleration', '接续需要二次采样加速'],+ summary: ['Turn on "Two-pass acceleration" and turn off the first-pass preview.', '请打开“二次采样加速”,并关闭首轮预览。']},{key: 'continuation-refine', message: 'Continue from video cannot be used with these settings. Clear either "Second pass of preview" or "Continue from video".',title: ['A continued part is made in full', '接续会完整生成这一段'],summary: ['Clear either "Second pass of preview" or "Continue from video".', '请清空“Second pass of preview”或“Continue from video”其中一项。']},diff --git a/web/freevideo.js b/web/freevideo.jsindex cfc33d1..cc15a3e 100644--- a/web/freevideo.js+++ b/web/freevideo.js@@ -686,15 +686,15 @@ app.registerExtension({if (prompt) { const owner = this; prompt.serializeValue = () => expandCast(prompt.value, mediaRows(owner), {active: !owner.freevideoCastOff?.()}).text; }const quality = this.widgets?.find(w => w.name === 'two_pass');if (quality) {- quality.label = text('Two-step generation', '两步生成');+ quality.label = text('Two-pass acceleration', '二次采样加速');quality.tooltip = text('Usually faster: generates at a lower resolution, then upscales and refines.', '通常更快:先用低分辨率生成,再放大细化。');}- for (const [name, en, zh] of [['base_steps', 'First-pass steps', '首轮步数'], ['refine_steps', 'Refine steps', '细化步数']]) {+ for (const [name, en, zh] of [['base_steps', 'First-pass steps', '一采步数'], ['refine_steps', 'Second-pass steps', '二采步数']]) {const steps = this.widgets?.find(w => w.name === name);if (steps) {steps.label = text(en, zh);- steps.tooltip = text('Changing steps may lower quality. Default: 8 first-pass and 3 refine steps. Two-step generation runs only with 8 first-pass steps. Other step counts use a single pass. Refine steps must be fewer than first-pass steps unless set to 3.',- '修改步数可能降低质量。默认首轮 8 步、细化 3 步;只有首轮为 8 步时才会使用两步生成,其他步数只生成一轮;细化步数不是 3 时,需少于首轮。');+ steps.tooltip = text('Changing sampling steps may lower generation quality. Default: 8 first-pass and 3 second-pass steps. Two-pass acceleration runs only with 8 first-pass steps. Other step counts use a single pass. Other second-pass counts must be fewer than first-pass steps.',+ '修改采样步数可能降低生成质量。默认一采 8 步、二采 3 步;只有一采为 8 步时才会使用二次采样加速,其他步数只采样一遍;其他二采步数需小于一采步数。');}}for (const widget of this.widgets || []) {diff --git a/web/generation_progress.js b/web/generation_progress.jsindex eea1a40..351a232 100644--- a/web/generation_progress.js+++ b/web/generation_progress.js@@ -35,7 +35,7 @@ const SAMPLING_NAME = {pass2: ['Generating video', '正在生成视频', 'Step', '正在生成 · 第 '],preview: ['Generating a preview', '正在生成预览', 'Step', '正在生成 · 第 '],};-const SAMPLING_PASS = {first: ['first pass', '首轮'], pass2: ['refine', '细化'], preview: ['preview', '预览']};+const SAMPLING_PASS = {first: ['first pass', '一采'], pass2: ['second pass', '二采'], preview: ['preview', '预览']};const SUBACTIVITY_LABELS = {'Starting generation process': 'start', 'Starting text encoder': 'libraries','Loading text encoder': 'libraries', 'Preparing text encoder GPU': 'gpu',@@ -180,17 +180,17 @@ function activityLabel(message, t, plan = {}, secondPass = false, references = fif (message.prediction || message.sampling_plan) {if (!ranOnePass(plan)) return named('prepare');const n = plan.base_steps;- return {key: 'prepare', text: t(`Two-step generation needs 8 first-pass steps. This video uses a single ${n}-step pass.`,- `只有首轮为 8 步时才会使用两步生成,这次按 ${n} 步只生成一轮。`)};+ return {key: 'prepare', text: t(`Two-pass acceleration needs 8 first-pass steps. This video uses a single ${n}-step pass.`,+ `只有一采为 8 步时才会使用二次采样加速,这次按 ${n} 步只采样一遍。`)};}if (phase === 'recovery') return message.retry?.reuse_sampling === true ? named('decode'): message.retry?.reuse_first_pass === true ? steps('pass2', 'pass2') : named('prepare');if (phase === 'sample_finalize') return named('decode');if (message.stage === 'latent_upscale' || phase === 'latent_upscale')return {key: 'upscale', text: Number.isInteger(second?.width) && Number.isInteger(second?.height)- ? t(`Refine · upscaling to ${second.width} × ${second.height}`,- `细化 · 放大到 ${second.width} × ${second.height}`)- : t('Refine · upscaling', '细化 · 放大')};+ ? t(`Second pass · upscaling to ${second.width} × ${second.height}`,+ `二采 · 放大到 ${second.width} × ${second.height}`)+ : t('Second pass · upscaling', '二采 · 放大')};if (phase === 'sampling' || phase === 'refinement' || message.stage === 'preview_mismatch') {const done = Number.isInteger(message.done) ? message.done : null;const total = Number.isInteger(message.total) && message.total > 0 ? message.total : null;@@ -205,7 +205,7 @@ function activityLabel(message, t, plan = {}, secondPass = false, references = fif (twoPass && plan.preview) return steps('preview', 'preview', step, count);const sampling = steps('pass1', twoPass ? 'first' : 'single', step, count);if (!twoPass && ranOnePass(plan))- sampling.text += t(' (single pass; two-step generation needs 8 first-pass steps)', '(只生成一轮;两步生成需首轮 8 步)');+ sampling.text += t(' (single pass; two-pass acceleration needs 8 first-pass steps)', '(只采样一遍;二次采样加速需一采 8 步)');return sampling;}const key = labelActivities.get(message.label) || {@@ -213,7 +213,7 @@ function activityLabel(message, t, plan = {}, secondPass = false, references = fload: 'load', decode: 'decode', save: 'save',}[phase] || labelActivities.get(message.stage);if (!key) return null;- if (key === 'refine') return {key: 'upscale', text: t('Refine · upscaling', '细化 · 放大')};+ if (key === 'refine') return {key: 'upscale', text: t('Second pass · upscaling', '二采 · 放大')};// Named encoder and post-sampling events stay distinct, even before a pause.const substage = SUBACTIVITY_LABELS[message.label] || (key === 'reference' ? 'reference' : null);if (substage) return {key: ['prompt', 'combined', 'reference'].includes(key) ? 'encode' : key,diff --git a/web/studio.js b/web/studio.jsindex 76cdd1b..cc40722 100644--- a/web/studio.js+++ b/web/studio.js@@ -631,7 +631,7 @@ export function openStudio(node, {extendFrom = null} = {}) {const twoPassStatus = el('small', '', 'fv-two-pass-status');// Off, on, or on and paused after the first pass: a preview is one way to run two-pass.const passMode = el('div', null, 'fv-pass-mode'); passMode.setAttribute('role', 'radiogroup');- passMode.setAttribute('aria-label', t('Two-step', '两步生成'));+ passMode.setAttribute('aria-label', t('Two-pass sampling', '二次采样'));const passOptions = [['off', t('Off', '关闭')], ['on', t('On', '开启')], ['preview', t('Preview first', '先出预览')]].map(([mode, label]) => {const option = el('button', label); option.type = 'button'; option.dataset.mode = mode; option.setAttribute('role', 'radio');option.onclick = () => choosePassMode(mode);@@ -650,7 +650,7 @@ export function openStudio(node, {extendFrom = null} = {}) {const next = enabled[(current + step + enabled.length) % enabled.length];if (next) { next.click(); next.focus(); }});- twoPassLabel.append(el('span', t('Two-step', '两步生成')), twoPassStatus, passMode);+ twoPassLabel.append(el('span', t('Two-pass', '二次采样')), twoPassStatus, passMode);let effortEstimate = {}, effortEstimateKey = '', effortEstimateTimer = null, effortEstimateRequest = null, lengthControl = null, chainController = null;const effort = createSamplingEffort(t, {onChange: tier => {baseSteps.value = tier.steps;@@ -745,8 +745,8 @@ export function openStudio(node, {extendFrom = null} = {}) {samplingWarning.id = `fv-sampling-warning-${node.id}`;samplingWarning.style.color = 'var(--fv-warning)';for (const [input, name, label, fallback, maximum] of [- [baseSteps, 'base_steps', t('First-pass steps', '首轮步数'), 8, 32],- [refineSteps, 'refine_steps', t('Refine steps', '细化步数'), 3, 31],+ [baseSteps, 'base_steps', t('First-pass steps', '一采步数'), 8, 32],+ [refineSteps, 'refine_steps', t('Second-pass steps', '二采步数'), 3, 31],]) {input.type = 'number'; input.min = '1'; input.max = String(maximum); input.step = '1'; input.required = true;input.value = value(node, name) ?? fallback;@@ -790,9 +790,9 @@ export function openStudio(node, {extendFrom = null} = {}) {baseSteps.min = twoPass.checked ? '2' : '1';refineSteps.max = baseLinked ? '31' : String(Math.max(3, Math.min(31, Number(baseSteps.value || 8) - 1)));refineSteps.setCustomValidity(twoPass.checked && !baseLinked && !refineSteps.disabled && Number(refineSteps.value) !== 3 && Number(refineSteps.value) >= Number(baseSteps.value)- ? t('Use 3 refine steps, or fewer steps than in the first pass.', '细化步数须为 3,或少于首轮步数。') : '');+ ? t('Use 3 second-pass steps, or fewer steps than in the first pass.', '二采步数须为 3,或少于一采步数。') : '');if (twoPass.checked) refineSteps.removeAttribute('title');- else refineSteps.title = t('Available when two-step is on.', '开启两步生成后可用。');+ else refineSteps.title = t('Turn on two-pass acceleration to use this setting.', '开启二次采样加速后生效。');for (const [input, name] of [[baseSteps, 'base_steps'], [refineSteps, 'refine_steps']]) {if (!input.disabled && input.checkValidity() && Number(input.value) !== value(node, name)) {set(node, name, Number(input.value));@@ -1338,8 +1338,8 @@ export function openStudio(node, {extendFrom = null} = {}) {const message = new Map([[seconds, t('Fill in the duration.', '请填写时长。')],[width, t('Fill in the width.', '请填写宽度。')], [height, t('Fill in the height.', '请填写高度。')],[seed, t('Fill in the seed.', '请填写种子。')],- [baseSteps, t('Fill in the first-pass steps.', '请填写首轮步数。')],- [refineSteps, t('Fill in the refine steps.', '请填写细化步数。')]]).get(invalid);+ [baseSteps, t('Fill in the first-pass steps.', '请填写一采步数。')],+ [refineSteps, t('Fill in the second-pass steps.', '请填写二采步数。')]]).get(invalid);if (message) { status.dataset.error = 'true'; status.textContent = message; }}invalid.focus(); invalid.reportValidity(); return;