With these Workflows (you'll also need to download my Custom Nodes via the ComfyUI Manager) you can create videos of pretty much infinite Length using the Minimax H3 First Frame Last Frame Model (see attached video example). The main idea is to use Multiple Keyframes to guide the video generation - but it should also work without them as long as you provide an initial starting frame. I'd recommend using Keyframes because they also serve as a 'quality reset' - in essence they prevent the typical Context and Quality loss that you normally see in long form AI video generation.<h3 id="included-workflows">Included workflows</h3>The following Workflows are included:1. Start Video — <code>Herrgotts_H3_Infinite_v1.2_01_Start.json</code> Use this Workflows to create your first Clip by providing a Start Frame, Prompt and end Frame. I've mainly tested 10 second generations, but it should theoretically work with any length. I wouldn't go for 5 seconds, since there's not enough time for the model to do it's work. After Clip generation, the Workflow and Custom Node automatically saves the latent that we will then use to continue the video.2. Continue Video into infinity — <code>Herrgotts_H3_Infinite_v1.2_02_Continue.json</code> Use this Workflows to continue any clip after the first one. Load the respective saved latent from the last Clip using the Custom Node, Input Prompt and (optionally) a new End Frame to guide your video. After the new Clip finished generating, the Workflow will, again, automatically save the new latent that you can then use to expand the video further using this Workflow. Just make sure that the correct latent Clip Index is loaded and saved according to the Output Folder. Using the same Clip Index of an existing latent will override that latent - so be careful!3. 3-Clip Showcase / Auto Stitch: Create one long Video entirely in one Workflow (similar to Wan SVI) — <code>Herrgotts_H3_Infinite_v1.2_03_3Clip_Showcase_AutoStitch.json</code> Runs Start Video -&gt; Continue Video -&gt; Continue Video in one queue and automatically creates a stitched final video. The graph is intentionally structured so another continuation block can be added for Clip 4+. You'll just have to provide one Starting Frame, All Prompts for the individual Clips and (optionally) End Frames for the Clips.4. Stitch Saved Chain — <code>Herrgotts_H3_Infinite_v1.2_04_Stitch_Saved_Chain.json</code> Combines clips generated separately with Workflows 1/2 and/ or 3. It decodes one saved AV latent at a time and writes directly to MP4, so memory usage does not grow with every clip in the chain. You won't have to worry about running OOM, no matter the video length.<h2 id="installation">Installation</h2>To install the Nodepack, just open one of the Workflows and use "Missing Custom Nodes" or Search fer "Herrgott" in ComfyUI Manager.<h2 id="main-features">Main features</h2><ul><li>Easy to use Workflows with intentionally simple structure so you can rearrange/ add nodes/ make adjustments easily.</li><li>Direct video + audio latent continuation without decode/re-encode handover (prevents quality loss and ensures smooth continuous motion).</li><li>Repeated Last Frame keyframe anchors for visual control and quality resets.</li><li>Auto Handover that detects the frozen FL2VA tail instead of using a fixed trim (H3 First Frame Last Frame often Produces Freeze Frames of the End Frame at the end of individual Clips, this prevents this from appearing in your videos).</li><li>Phase-aligned context to avoid startup flicker from invalid H3 latent cut positions.</li><li>No-Lock Fallback when no final freeze is detected (added this to make sure the video quality remains consistent during the whole video).</li><li>Safe Tail Bridge: rendered frames lost only because of latent phase alignment can replace the first 1–2 potentially unstable video frames of the next clip (prevents small lumination changes between Clips from mattering as much).</li><li>Short context-aligned video crossfade and independently tested 15 ms audio de-click crossfade.</li><li><code>Full</code>, <code>Stitch Ready</code> and <code>Final Clip</code> output modes.</li><li>Save / Load complete AV latents with stitch metadata.</li><li>A memory-bounded Saved Chain Stitcher for long projects generated clip by clip.</li><li>Lazy, marker-gated H3 runtime hooks that are installed only when continuation is actually used.</li></ul>If you use the Nodepack or the Workflows, it'd be awesome if you could share your favorite creations! Have fun Prompting. :)

H3 Infinite Video Continuation and automatic Stitching with Fl2Va Model

v1.3.0 adds flexible native MiniMax H3 conditioning and Qwen image guides while keeping the proven v1.2 continuation and stitching core intact.HighlightsNew Flexible Start / Conditioning v1.3 node.<ul><li>First Frame and Last Frame are independent optional inputs.</li><li>Supports T2VA, I2VA, L2VA and FL2VA from one node.</li></ul>New Continue from Latent v1.3 node.<ul><li>Keeps direct phase-aligned video + audio latent continuation.</li><li>Adds optional Last Frame conditioning for a fresh visual endpoint / quality reset.</li></ul>Added auto-growing Qwen Reference inputs.<ul><li>Starts with <code>Qwen Reference 1</code> and grows up to 9 references as sockets are connected.</li><li>Qwen References are text/vision-encoder-only images and are not inserted as persistent native Ref2VA/DiT reference latents.</li></ul>Added deterministic <code>&lt;Picture N&gt;</code> mapping and a <code>picture_map</code> diagnostic output.<ul><li>Connected First Frame comes first.</li><li>Connected Last Frame comes next.</li><li>Qwen References follow in numeric order.</li><li>Direct continuation latent context is not silently counted as a Qwen Picture.</li></ul><ul><li>Existing v1.2.x nodes remain registered for workflow compatibility.</li><li>Updated all four example workflows and Registry metadata to v1.3.0.</li></ul>

Files with SHA256: 72c562530cc9babdd93fd9cf3daafa4d34b9a2d616dfa443ae61acf165228688