Fix FlashVSR ghosting: streaming TCDecoder decode + Causal LQ projection

Root cause: three critical differences from naxci1 reference implementation:

1. Batch decode after loop → streaming per-chunk TCDecoder decode with LQ
   conditioning inside the loop. The TCDecoder uses causal convolutions with
   temporal memory that must be built incrementally per-chunk. Batch decode
   breaks this design and loses LQ frame conditioning, causing ghosting.

2. Buffer_LQ4x_Proj → Causal_LQ4x_Proj for FlashVSR v1.1. The causal
   variant reads the OLD cache before writing the new one (truly causal),
   while Buffer writes cache BEFORE the conv call. Using the wrong variant
   misaligns temporal LQ conditioning features.

3. Temporal padding formula: changed from round-up to largest_8n1_leq(N+4)
   matching the naxci1 reference approach.

Changes:
- flashvsr_full.py: streaming TCDecoder decode per-chunk with LQ conditioning
  and per-chunk color correction (was: batch VAE decode after loop)
- flashvsr_tiny.py: streaming TCDecoder decode per-chunk (was: batch decode)
- inference.py: use Causal_LQ4x_Proj, build TCDecoder for ALL modes (including
  full), fix temporal padding to largest_8n1_leq(N+4), clear TCDecoder in
  clear_caches()
- utils.py: add Causal_LQ4x_Proj class
- nodes.py: update progress bar estimation for new padding formula

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

flashvsr_arch/models/utils.py

@@ -357,4 +357,104 @@ class Buffer_LQ4x_Proj(nn.Module):
                 outputs.append(self.linear_layers[i](out_x))
             self.clip_idx += 1
             return outputs
-        
+
+
+class Causal_LQ4x_Proj(nn.Module):
+    """Causal variant of Buffer_LQ4x_Proj for FlashVSR v1.1.
+
+    Key difference: reads old cache BEFORE writing new cache (truly causal),
+    whereas Buffer_LQ4x_Proj writes cache BEFORE conv call.
+    """
+
+    def __init__(self, in_dim, out_dim, layer_num=30):
+        super().__init__()
+        self.ff = 1
+        self.hh = 16
+        self.ww = 16
+        self.hidden_dim1 = 2048
+        self.hidden_dim2 = 3072
+        self.layer_num = layer_num
+
+        self.pixel_shuffle = PixelShuffle3d(self.ff, self.hh, self.ww)
+
+        self.conv1 = CausalConv3d(in_dim*self.ff*self.hh*self.ww, self.hidden_dim1, (4, 3, 3), stride=(2, 1, 1), padding=(1, 1, 1))
+        self.norm1 = RMS_norm(self.hidden_dim1, images=False)
+        self.act1 = nn.SiLU()
+
+        self.conv2 = CausalConv3d(self.hidden_dim1, self.hidden_dim2, (4, 3, 3), stride=(2, 1, 1), padding=(1, 1, 1))
+        self.norm2 = RMS_norm(self.hidden_dim2, images=False)
+        self.act2 = nn.SiLU()
+
+        self.linear_layers = nn.ModuleList([nn.Linear(self.hidden_dim2, out_dim) for _ in range(layer_num)])
+
+        self.clip_idx = 0
+
+    def forward(self, video):
+        self.clear_cache()
+        t = video.shape[2]
+        iter_ = 1 + (t - 1) // 4
+        first_frame = video[:, :, :1, :, :].repeat(1, 1, 3, 1, 1)
+        video = torch.cat([first_frame, video], dim=2)
+
+        out_x = []
+        for i in range(iter_):
+            x = self.pixel_shuffle(video[:, :, i*4:(i+1)*4, :, :])
+            cache1_x = x[:, :, -CACHE_T:, :, :].clone()
+            x = self.conv1(x, self.cache['conv1'])       # reads OLD cache
+            self.cache['conv1'] = cache1_x               # writes NEW cache AFTER
+            x = self.norm1(x)
+            x = self.act1(x)
+            cache2_x = x[:, :, -CACHE_T:, :, :].clone()
+            if i == 0:
+                self.cache['conv2'] = cache2_x
+                continue
+            x = self.conv2(x, self.cache['conv2'])       # reads OLD cache
+            self.cache['conv2'] = cache2_x               # writes NEW cache AFTER
+            x = self.norm2(x)
+            x = self.act2(x)
+            out_x.append(x)
+        out_x = torch.cat(out_x, dim=2)
+        out_x = rearrange(out_x, 'b c f h w -> b (f h w) c')
+        outputs = []
+        for i in range(self.layer_num):
+            outputs.append(self.linear_layers[i](out_x))
+        return outputs
+
+    def clear_cache(self):
+        self.cache = {}
+        self.cache['conv1'] = None
+        self.cache['conv2'] = None
+        self.clip_idx = 0
+
+    def stream_forward(self, video_clip):
+        if self.clip_idx == 0:
+            first_frame = video_clip[:, :, :1, :, :].repeat(1, 1, 3, 1, 1)
+            video_clip = torch.cat([first_frame, video_clip], dim=2)
+            x = self.pixel_shuffle(video_clip)
+            cache1_x = x[:, :, -CACHE_T:, :, :].clone()
+            x = self.conv1(x, self.cache['conv1'])       # reads OLD (None) cache
+            self.cache['conv1'] = cache1_x               # writes AFTER
+            x = self.norm1(x)
+            x = self.act1(x)
+            cache2_x = x[:, :, -CACHE_T:, :, :].clone()
+            self.cache['conv2'] = cache2_x
+            self.clip_idx += 1
+            return None
+        else:
+            x = self.pixel_shuffle(video_clip)
+            cache1_x = x[:, :, -CACHE_T:, :, :].clone()
+            x = self.conv1(x, self.cache['conv1'])       # reads OLD cache
+            self.cache['conv1'] = cache1_x               # writes AFTER
+            x = self.norm1(x)
+            x = self.act1(x)
+            cache2_x = x[:, :, -CACHE_T:, :, :].clone()
+            x = self.conv2(x, self.cache['conv2'])       # reads OLD cache
+            self.cache['conv2'] = cache2_x               # writes AFTER
+            x = self.norm2(x)
+            x = self.act2(x)
+            out_x = rearrange(x, 'b c f h w -> b (f h w) c')
+            outputs = []
+            for i in range(self.layer_num):
+                outputs.append(self.linear_layers[i](out_x))
+            self.clip_idx += 1
+            return outputs