Upload SVIPro_wan22_v1.json

SVIPro_wan22_v1.json

@@ -5008,54 +5008,6 @@
       "color": "#223",
       "bgcolor": "#335"
     },
-    {
-      "id": 319,
-      "type": "MarkdownNote",
-      "pos": [
-        5330,
-        1690
-      ],
-      "size": [
-        500,
-        790
-      ],
-      "flags": {},
-      "order": 77,
-      "mode": 0,
-      "inputs": [],
-      "outputs": [],
-      "title": "README -  WAN 2.2 general guide",
-      "properties": {},
-      "widgets_values": [
-        "## This workflow has a few specialities:  \n  \n1) The Lightx2v distill loras are used during during both high and low noise passes, which drastically increases generation speed at the cost of less dynamic/detailed generations. Distill model/lora usage requires CFG 1.0, and works within 4-8 total steps. You should distribute the steps evenly across both phases.\n    \n2) The raw (16fps) video is not saved by default, but the frames are run through VFI to double the frame rate to 32fps, giving a much smoother playback. If you don't want that, enable \"save_output\" on the first Video Combine node and bypass VFI and the second one.  \n  \n3) In case you want to try to extend the video, the raw frames of either the 16fps or VFI video can be saved alongside the final video file. Enable/disable one of the Save Image nodes accordingly. It is recommended to use the RAW frames as those adhere to the Wan 16fps standard.\n  \n4) The video file and frames are by default saved into a subfolder that is named after the seed used by the high noise ksampler.  \n\n**Do NOT** use the distill loras together with a model that has them merged in! Read the model description to verify if that is the case or not.  \n\n**Shift**  \nUse 5.0 for t2v and 8.0 for i2v. Values within that range usually work well.\n  \n**Width & Height**  \nWan 2.2 is trained on ~1 megapixel, the default resolutions are  \n1280 x 720  \n720 x 1280  \n960 x 960  \nbut in general it is very flexible about the resolution and ratio so you can use lower values if you run out of vram.\n\n**Length**  \nDefault length of Wan 2.2 videos is 81 frames which corresponds to ~5 seconds of video. Usually any range between 61 and 121 frames works okay, anything further off results in incoherent output.  \n\n## Dual KSampler setup (MoE model)\nWan 2.2. is a Mixture of Experts model, which means it has one model that creates the basic video composition and motion from the empty latent, while a second model then refines its output to increase details and quality. This requires a specific setup:  \n  \n1) **Both** KSamplers must be set to the same total amount of steps!\n2) The first KSampler \"end_at_step\" should be at half the amount of total steps and \"return_leftover_noise\" must be enabled.\n3) The second KSampler \"start_at_step\" should be the same value as the first ones \"end_at_step\", while its own \"end_at_step\" must be equal to the total amount of steps.\n\n## _Distill parameter setup_\n  \n**Total steps**: minimum of 4, maximum of 8\n\n**CFG**: The distill loras require CFG to be 1.0.  \nThis disables the negative prompt!  \n  \n## _Normal parameter setup_\n\n(Bypass distill lora loader nodes)  \n  \n**Total steps**: ~20\n\n**CFG**: ~3.5  \nThis enables the negative prompt at the cost of increasing generation time x1.5, but is required for coherent results.\n"
-      ],
-      "color": "#332922",
-      "bgcolor": "#593930"
-    },
-    {
-      "id": 324,
-      "type": "MarkdownNote",
-      "pos": [
-        5860,
-        1690
-      ],
-      "size": [
-        450,
-        790
-      ],
-      "flags": {},
-      "order": 78,
-      "mode": 0,
-      "inputs": [],
-      "outputs": [],
-      "title": "README - SVI 2.0 Pro",
-      "properties": {},
-      "widgets_values": [
-        "## How it works\n\nSVI uses the input image and the last 8 frames of the previous extent to get some context for the next clip. This is an improvement over normal i2v as it will have some basic idea about e.g. speed and direction of things from the previous clip, as well as it can keep characters more consistent along the extents with the information of the input image.\n\nYou can use any WAN 2.2. dual-model merge with SVI, so enable/disable the distil lora load nodes accordingly.\n\nYou can enable/disable or copy and paste more extent subgraphs and the corresponding get nodes to make longer/shorter videos. In theory there is no limit how many you line up, but expect some drift over time.\n\n## Extents\nCurrently the extents are optimized for distil mode (4 steps cfg 1), and that setting is not exposed on the subgraph node, so if you want to do it differently you have to change it in all of them!\n\n## \"Start Prompt\"\nThis is prepended to each extents individual prompt.\n\nThis should describe some overall common scene/char setup, but not any action. If you want to have drastic changes on your extents, don't describe anything here that would disappear later, in that case rather leave it empty and just use the individual prompt boxes\n\n## Prompts\nStart each prompt with a short description of how the previous clip ended. You don't need to include specific actions, but explaining the setup of the last frame can help WAN understand it better. SVI is still a hack to extent videos, so the more help you give WAN the better.\n\nDon't overprompt each clip. There is only so much things and details WAN can do in one clip, if you want to have a lot of things to happen it can be beneficial to make shorter \"transition\" clips with less frames.\n\n## Tip\nGenerating all extents at once is a gamble cause if one of them is bad you can throw away the full clip. The more reliable process is to roll until you get a good first clip, then set that seed to \"fixed\".\n\n**It is highly recommended to set \"Widget control mode\" to \"before\" in the ComfyUI settings, that will leave the seed that was used for that last run in the RandomNoise nodes!**\n\nWhen you set a subgraphs seed to fixed, Comfy will not regenerate it but jump directly to the next one when you Run again. This way you can roll each consecutive clip until its good, then rinse repeat with the next one."
-      ],
-      "color": "#432",
-      "bgcolor": "#653"
-    },
     {
       "id": 196,
       "type": "VHS_VideoCombine",
@@ -5244,6 +5196,54 @@
       ],
       "color": "#223",
       "bgcolor": "#335"
+    },
+    {
+      "id": 319,
+      "type": "MarkdownNote",
+      "pos": [
+        5330,
+        1690
+      ],
+      "size": [
+        500,
+        800
+      ],
+      "flags": {},
+      "order": 77,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [],
+      "title": "README -  WAN 2.2 general guide",
+      "properties": {},
+      "widgets_values": [
+        "## This workflow has a few specialities:  \n  \n1) The Lightx2v distill loras are used during during both high and low noise passes, which drastically increases generation speed at the cost of less dynamic/detailed generations. Distill model/lora usage requires CFG 1.0, and works within 4-8 total steps. You should distribute the steps evenly across both phases.\n    \n2) The raw (16fps) video is not saved by default, but the frames are run through VFI to double the frame rate to 32fps, giving a much smoother playback. If you don't want that, enable \"save_output\" on the first Video Combine node and bypass VFI and the second one.  \n  \n3) In case you want to try to extend the video, the raw frames of either the 16fps or VFI video can be saved alongside the final video file. Enable/disable one of the Save Image nodes accordingly. It is recommended to use the RAW frames as those adhere to the Wan 16fps standard.\n  \n4) The video file and frames are by default saved into a subfolder that is named after the seed used by the high noise ksampler.  \n\n**Do NOT** use the distill loras together with a model that has them merged in! Read the model description to verify if that is the case or not.  \n\n**Shift**  \nUse 5.0 for t2v and 8.0 for i2v. Values within that range usually work well.\n  \n**Width & Height**  \nWan 2.2 is trained on ~1 megapixel, the default resolutions are  \n1280 x 720  \n720 x 1280  \n960 x 960  \nbut in general it is very flexible about the resolution and ratio so you can use lower values if you run out of vram.\n\n**Length**  \nDefault length of Wan 2.2 videos is 81 frames which corresponds to ~5 seconds of video. Usually any range between 61 and 121 frames works okay, anything further off results in incoherent output.  \n\n## Dual KSampler setup (MoE model)\nWan 2.2. is a Mixture of Experts model, which means it has one model that creates the basic video composition and motion from the empty latent, while a second model then refines its output to increase details and quality. This requires a specific setup:  \n  \n1) **Both** KSamplers must be set to the same total amount of steps!\n2) The first KSampler \"end_at_step\" should be at half the amount of total steps and \"return_leftover_noise\" must be enabled.\n3) The second KSampler \"start_at_step\" should be the same value as the first ones \"end_at_step\", while its own \"end_at_step\" must be equal to the total amount of steps.\n\n## _Distill parameter setup_\n  \n**Total steps**: minimum of 4, maximum of 8\n\n**CFG**: The distill loras require CFG to be 1.0.  \nThis disables the negative prompt!  \n  \n## _Normal parameter setup_\n\n(Bypass distill lora loader nodes)  \n  \n**Total steps**: ~20\n\n**CFG**: ~3.5  \nThis enables the negative prompt at the cost of increasing generation time x1.5, but is required for coherent results.\n\n[Xeno443](https://huggingface.co/datasets/Xeno443/ComfyUI_Workflows/tree/main)\n"
+      ],
+      "color": "#332922",
+      "bgcolor": "#593930"
+    },
+    {
+      "id": 324,
+      "type": "MarkdownNote",
+      "pos": [
+        5860,
+        1690
+      ],
+      "size": [
+        450,
+        800
+      ],
+      "flags": {},
+      "order": 78,
+      "mode": 0,
+      "inputs": [],
+      "outputs": [],
+      "title": "README - SVI 2.0 Pro",
+      "properties": {},
+      "widgets_values": [
+        "## How it works\n\nSVI uses the input image and the last 8 frames of the previous extent to get some context for the next clip. This is an improvement over normal i2v as it will have some basic idea about e.g. speed and direction of things from the previous clip, as well as it can keep characters more consistent along the extents with the information of the input image.\n\nYou can use any WAN 2.2. dual-model merge with SVI, so enable/disable the distil lora load nodes accordingly.\n\nYou can enable/disable or copy and paste more extent subgraphs and the corresponding get nodes to make longer/shorter videos. In theory there is no limit how many you line up, but expect some drift over time.\n\n## Extents\nCurrently the extents are optimized for distil mode (4 steps cfg 1), and that setting is not exposed on the subgraph node, so if you want to do it differently you have to change it in all of them!\n\n## \"Start Prompt\"\nThis is prepended to each extents individual prompt.\n\nThis should describe some overall common scene/char setup, but not any action. If you want to have drastic changes on your extents, don't describe anything here that would disappear later, in that case rather leave it empty and just use the individual prompt boxes\n\n## Prompts\nStart each prompt with a short description of how the previous clip ended. You don't need to include specific actions, but explaining the setup of the last frame can help WAN understand it better. SVI is still a hack to extent videos, so the more help you give WAN the better.\n\nDon't overprompt each clip. There is only so much things and details WAN can do in one clip, if you want to have a lot of things to happen it can be beneficial to make shorter \"transition\" clips with less frames.\n\n## Tip\nGenerating all extents at once is a gamble cause if one of them is bad you can throw away the full clip. The more reliable process is to roll until you get a good first clip, then set that seed to \"fixed\".\n\n**It is highly recommended to set \"Widget control mode\" to \"before\" in the ComfyUI settings, that will leave the seed that was used for that last run in the RandomNoise nodes!**\n\nWhen you set a subgraphs seed to fixed, Comfy will not regenerate it but jump directly to the next one when you Run again. This way you can roll each consecutive clip until its good, then rinse repeat with the next one."
+      ],
+      "color": "#432",
+      "bgcolor": "#653"
     }
   ],
   "links": [