FreeVideo 0.4.7

README.zh-CN.md

@@ -20,7 +20,7 @@ https://github.com/user-attachments/assets/ecda7d0d-7fbe-4e0c-8c29-8f3315bafc15
 ## 更新动态
 
 - **2026-10-13** · **[v0.4.0](https://github.com/FlashML-org/FreeVideo/releases/tag/v0.4.0)：Agent 全自动生成、分钟级长视频和剪辑视图。** 连接 Claude Code、Codex 等 AI Agent，让它全自动生成视频；一条视频可以生成到一分钟以上；在时间线上剪辑。
-- **2026-10-07** · **[v0.3.2](https://github.com/FlashML-org/FreeVideo/releases/tag/v0.3.2)：先出预览。** 先看半分辨率预览，满意的再继续二采。
+- **2026-10-07** · **[v0.3.2](https://github.com/FlashML-org/FreeVideo/releases/tag/v0.3.2)：先出预览。** 先看半分辨率预览，满意的再继续生成。
 - **2026-10-07** · **[v0.3.0](https://github.com/FlashML-org/FreeVideo/releases/tag/v0.3.0)：更快的 int8 模型。** GeForce 和 RTX 30 系显卡改用 int8，RTX 3090 生成 5 秒视频快约 2.3 倍。
 - **2026-10-06** · **[作品展示](https://freevideo-community.pages.dev/#gallery)上线。** 观看 FreeVideo 生成的 20 秒视频，以及四档质量的并排对比。
 - **2026-10-06** · **[v0.2.3](https://github.com/FlashML-org/FreeVideo/releases/tag/v0.2.3)：视频自带工作流。** 把 FreeVideo 生成的视频拖到 ComfyUI 画布上，即可还原提示词、种子和全部参数，之前的视频也能补上。

docs/Linux.md

@@ -48,7 +48,7 @@ To connect from computers on your local network without SSH:
 freevideo server --listen 0.0.0.0
 ```
 
-This prints an access link with a token. Anyone with the link can use FreeVideo on the server; `freevideo server --listen 0.0.0.0 --new-token` replaces it, and `freevideo status` shows it again. ComfyUI itself still listens only on the server; a small proxy in front of it turns away every request without the token. The connection is not encrypted, so use it on a network you trust, or add `--tls-cert cert.pem --tls-key key.pem` for HTTPS.
+This prints an access link with a token, and a QR code of it to scan with a phone camera. Anyone with the link can use FreeVideo on the server; `freevideo server --listen 0.0.0.0 --new-token` replaces it, and `freevideo status` shows it again. ComfyUI itself still listens only on the server; a small proxy in front of it turns away every request without the token. The connection is not encrypted, so use it on a network you trust, or add `--tls-cert cert.pem --tls-key key.pem` for HTTPS.
 
 To start FreeVideo with the server, add `--enable`; it becomes a systemd user service. To keep it running after you log out and start it at boot, also run `sudo loginctl enable-linger $USER` once. `freevideo server --disable` removes the service.
 

docs/Linux.zh-CN.md

@@ -48,7 +48,7 @@ ssh -N -L 8188:127.0.0.1:8188 用户名@服务器
 freevideo server --listen 0.0.0.0
 ```
 
-这会打印一个带访问令牌的链接，持有链接的人都可以使用这台服务器上的 FreeVideo；用 `freevideo server --listen 0.0.0.0 --new-token` 可以换一个新链接，`freevideo status` 可以再次查看。ComfyUI 本身仍然只监听服务器本机，前面的访问代理会拒绝所有不带令牌的请求。连接没有加密，请只在可信的网络中使用，或加上 `--tls-cert cert.pem --tls-key key.pem` 使用 HTTPS。
+这会打印一个带访问令牌的链接和对应的二维码（可用手机相机扫描），持有链接的人都可以使用这台服务器上的 FreeVideo；用 `freevideo server --listen 0.0.0.0 --new-token` 可以换一个新链接，`freevideo status` 可以再次查看。ComfyUI 本身仍然只监听服务器本机，前面的访问代理会拒绝所有不带令牌的请求。连接没有加密，请只在可信的网络中使用，或加上 `--tls-cert cert.pem --tls-key key.pem` 使用 HTTPS。
 
 如需随服务器启动，加上 `--enable`，FreeVideo 会注册为 systemd 用户服务；若要在注销后保持运行、开机自动启动，还需运行一次 `sudo loginctl enable-linger $USER`。`freevideo server --disable` 会移除这个服务。
 

freevideo_engine/adaln.py

@@ -12,6 +12,20 @@
 import torch
 
 
+def all_finite(tensor):
+    """Whether every element is finite; reads half-precision bits instead of calling isfinite.
+
+    A bfloat16 or float16 table is finite exactly when its exponent field is not
+    all ones, so one integer mask answers the question about ten times faster on
+    the CPU than isfinite() on those dtypes.
+    """
+    mask = {torch.bfloat16: 0x7F80, torch.float16: 0x7C00}.get(tensor.dtype)
+    if mask is None:
+        return bool(torch.isfinite(tensor).all())
+    bits = tensor.reshape(-1).view(torch.int16)
+    return not bool((bits & mask).eq(mask).any())
+
+
 class ScheduleCursor:
     def __init__(self, timesteps):
         self.timesteps = timesteps
@@ -174,7 +188,7 @@ def load(self, index, steps):
         # These are small immutable constants; their producer's GPU/runtime is
         # deliberately irrelevant. Validate before transferring to this device.
         state = load_file(path, device=self.device)
-        if any(not bool(torch.isfinite(tensor).all()) for tensor in state.values()):
+        if any(not all_finite(tensor) for tensor in state.values()):
             raise ValueError('Nonfinite AdaLN model asset')
         return [state[f'step_{i}'].chunk(6, dim=-1) for i in range(steps)]
 
@@ -199,7 +213,7 @@ def save(self, index, module, *, producer_device=None):
         path = self.root / f'{index:02d}.safetensors'
         temporary = path.with_suffix('.partial')
         state = {name: tensor.cpu() for name, tensor in module.state_dict().items()}
-        if any(not bool(torch.isfinite(tensor).all()) for tensor in state.values()):
+        if any(not all_finite(tensor) for tensor in state.values()):
             raise ValueError('Cannot save nonfinite AdaLN constants')
         save_file(state, temporary)
         row = {'bytes': temporary.stat().st_size, 'sha256': assets.file_hash(temporary)}

freevideo_engine/agent/guide.py

@@ -70,6 +70,13 @@
         'lighting and surroundings as described above; do not copy its simplified, flat look." The result follows '
         'the draft and looks plainer than a direct generation, so keep the direct one when the user prefers its look.',
     ],
+    'quality': [
+        'Steps per quality: "light" = 8 first-pass steps + 3 refine steps, two-pass (default); "medium" = 12 steps, '
+        '"high" = 16 steps, "max" = 20 steps, each in one pass.',
+        'Two-pass generation runs only with 8 first-pass steps. Any other base_steps (for example 20 with '
+        'two_pass=true) runs in one pass of that many steps; the job notice says so. A preview is always two-pass '
+        'with 8 steps.',
+    ],
     'timing': [
         'Single-block frames are 24 fps on a 17n+5 grid (6 s -> 6.583 s, 10 s -> 10.125 s). '
         'Under 30 s defaults to one generation when the device supports this size and quality; unknown limits use 15 s. '

freevideo_engine/agent/jobs.py

@@ -131,8 +131,14 @@ def continuation_problems(spec, prefix='', *, editor=None):
     problems = []
     if not spec['two_pass']:
         # The quality name is a contract enum, never caller text.
-        problem = schema.diagnostic(prefix + '/quality', '"light" (two-pass), or sampling.two_pass=true,' + why, None)
-        got = '"%s" (one pass)' % spec['quality']
+        if spec.get('sampling_note'):
+            # Two-pass was asked for with a first-pass count other than 8: it runs one pass.
+            problem = schema.diagnostic(prefix + '/quality',
+                                        '"light" (two-pass), or sampling.two_pass=true with base_steps=8,' + why, None)
+            got = '%d first-pass steps (one pass)' % spec['base_steps']
+        else:
+            problem = schema.diagnostic(prefix + '/quality', '"light" (two-pass), or sampling.two_pass=true,' + why, None)
+            got = '"%s" (one pass)' % spec['quality']
         problems.append(dict(problem, got=got, message=problem['message'].replace('received None', 'received ' + got)))
     if spec['preview']:
         problems.append(schema.diagnostic(prefix + '/preview', 'false' + why, True))
@@ -546,6 +552,12 @@ def load(self):
                     self.finish(row, 'failed', error=dict(code='interrupted',
                                 message='ComfyUI restarted while this job ran; submit it again'), save=False)
                 row['_restored'] = True
+                if row['kind'] == 'generate' and row['status'] in ACTIVE:
+                    # Waiting, queued or running when the process restarted: this process runs it
+                    # under the one-pass rule, however it is started again. A finished job is left.
+                    note = self.legacy_one_pass_note(row['spec'])
+                    if note:
+                        row['spec']['sampling_note'] = note
                 if row['kind'] == 'generate' and not row['spec'].get('refine_from'):
                     # Before device-aware Agent planning, every accepted job used the unknown limit.
                     row['spec'].setdefault('single_max_frames', None)
@@ -1079,6 +1091,15 @@ def view(self, row, *, detail=True):
             result['chain_id'] = row['chain_id']
         if row['kind'] == 'generate':
             result.update(cancel_requested=row.get('cancel_requested', False), attempt=row.get('attempt', 1))
+            note = spec.get('sampling_note')
+            if (note is None and row['status'] not in DONE and spec.get('two_pass') and not spec.get('preview')
+                    and not spec.get('refine_from')):
+                # A job stored before the one-pass rule and not yet run: the engine applies it when it runs.
+                # A finished one keeps what it did, which a stored request cannot tell.
+                from ..two_pass import one_pass_note
+                note = one_pass_note(spec.get('base_steps', 8))
+            if note:
+                result['notice'] = note
         if row['status'] == 'waiting':
             pending = [d for d in row.get('depends_on', ()) if self.jobs.get(d, {}).get('status') not in DONE]
             result['waiting_for'] = pending or row.get('waiting_for') or (['gpu'] if row['kind'] in GPU_KINDS + ('generate',) else [])
@@ -1449,9 +1470,28 @@ def media_file(self, row, item, name, *, target=None):
             relative = edited
         return relative
 
+    @staticmethod
+    def legacy_one_pass_note(spec):
+        """The note for an unfinished two-pass job stored before the one-pass rule, else None."""
+        base = spec.get('base_steps', 8)
+        if (spec.get('sampling_note') is None and spec.get('two_pass') is True and not spec.get('preview')
+                and not spec.get('refine_from') and type(base) is int and 1 <= base <= 32 and base != 8):
+            from ..two_pass import one_pass_note
+            return one_pass_note(base)
+        return None  # a damaged stored value gets no note; the row is judged where it is used
+
     async def start_generation(self, row):
         def prepare():
             spec = row['spec']
+            note = self.legacy_one_pass_note(spec)
+            if note:
+                # A job stored before the one-pass rule starts under it now, on a first start
+                # and on a resume alike: keep what it runs as, so the note outlasts the run and
+                # a reload (a stored request alone cannot tell, and a finished old job must
+                # not gain a note).
+                with self.lock:
+                    spec['sampling_note'] = note
+                self.touch(row)
             if row.get('resume_pending'):
                 self.start_chain(row, resume=True)
                 return

freevideo_engine/agent/routes.py

@@ -251,6 +251,8 @@ def plan(port, value, jobs=None):
     result['geometry'] = {k: canvas[k] for k in ('width', 'height', 'frames', 'seconds', 'fps', 'latent_frames')}
     result['sampling'] = dict(two_pass=passes['enabled'], total_steps=passes['total_steps'],
                               first_pass=passes.get('first'), second_pass=passes.get('second') if passes['enabled'] else None)
+    if spec.get('sampling_note'):
+        result['warnings'].append(spec['sampling_note'])
     if spec['seconds'] > 15:
         result['warnings'].append('Longer than 15 s is experimental')
     if spec['task'] == 'ref2va':

freevideo_engine/agent/schema.py

@@ -176,7 +176,22 @@ def generate(*, local=False):
                {'if': {'if': {'required': ['sampling'], 'properties': {'sampling': {'required': ['two_pass']}}},
                        'then': {'properties': {'sampling': {'properties': {'two_pass': {'const': False}}}}},
                        'else': {'required': ['quality'], 'properties': {'quality': {'not': {'const': 'light'}}}}},
-                'then': {'properties': {'preview': {'const': False}}}}],
+                'then': {'properties': {'preview': {'const': False}}}},
+               # A preview is two-pass, which runs only with 8 first-pass steps: refused when the effective
+               # first-pass count is another number, given in sampling or inherited from the quality.
+               {'if': {'properties': {'preview': {'const': True}}, 'required': ['preview']},
+                # x-derived: issues() names this case at /sampling with the derived sampling checks.
+                'then': {'x-derived': True, 'not': {'anyOf': [
+                    {'required': ['sampling'], 'properties': {'sampling': {'required': ['two_pass', 'base_steps'],
+                        'properties': {'two_pass': {'const': True}, 'base_steps': {'not': {'const': 8}}}}}},
+                    {'required': ['sampling'], 'properties': {
+                        'sampling': {'required': ['base_steps'], 'not': {'required': ['two_pass']},
+                                     'properties': {'base_steps': {'not': {'const': 8}}}},
+                        'quality': {'const': 'light'}}},
+                    {'required': ['sampling', 'quality'], 'properties': {
+                        'sampling': {'required': ['two_pass'], 'not': {'required': ['base_steps']},
+                                     'properties': {'two_pass': {'const': True}}},
+                        'quality': {'enum': ['medium', 'high', 'max']}}}]}}}],
         description='x-job-kind selects normalization and checks derived geometry and sampling step comparisons; these calculations are not JSON Schema constraints.',
         **{'x-job-kind': 'generate'})
 
@@ -976,7 +991,7 @@ def bad(allowed, got=value, suffix='', **kw):
     if 'if' in schema:
         rule = schema.get('then', True) if not issues(schema['if'], value) else schema.get('else', True)
         errors += issues(rule, value, path)
-    if 'not' in schema and not issues(schema['not'], value):
+    if 'not' in schema and not schema.get('x-derived') and not issues(schema['not'], value):
         errors.append(diagnostic(path, 'a permitted field combination', value))
     if 'oneOf' in schema or 'anyOf' in schema:
         # Structural unions choose the most specific applicable branch, then
@@ -1101,6 +1116,8 @@ def bad(field, allowed, got=None):
                 validate_steps(base, refine, two)
             except ValueError:
                 bad('sampling', 'two-pass refine_steps=3 or refine_steps below base_steps', custom)
+            if value.get('preview') and two and base != 8:
+                bad('sampling', 'base_steps=8 for a preview, the only first-pass count that runs two-pass', custom)
     if kind == 'vlm':
         rows = value['questions'] if isinstance(value['questions'], list) else [value['questions']]
         count = sum(len(r.get('images', [])) + (r.get('video', {}).get('frames', 6) if isinstance(r.get('video'), dict)

freevideo_engine/agent/spec.py

@@ -130,9 +130,16 @@ def sampling(value):
     two, base, refine = QUALITY[quality]
     custom = value.get('sampling', {})
     two, base, refine = custom.get('two_pass', two), custom.get('base_steps', base), custom.get('refine_steps', refine)
+    note = None
+    if two:
+        from ..two_pass import one_pass_note
+        # Only the 8-step first pass goes two-pass; any other count is one pass, said once in the answer.
+        note = one_pass_note(base)
+        two = two and note is None
     if custom:
         quality = next((k for k, v in QUALITY.items() if v == (two, base, refine)), 'custom')
-    return dict(quality=quality, two_pass=two, base_steps=base, refine_steps=refine)
+    return dict(quality=quality, two_pass=two, base_steps=base, refine_steps=refine,
+                **({'sampling_note': note} if note else {}))
 
 
 def generation_media(value):