MiniMax H3 一键高清放大 - Star7:
MiniMaxH3OneClickHDStar7 enhances a generated MiniMax H3 video using a learned latent upscaler followed by an optional, short refinement pass. Instead of enlarging already decoded images, it works on the video’s compressed internal representation, allowing resolution enhancement before final decoding. The refinement stage revisits a controlled portion of the generation process to improve the enlarged result, with lower strength favoring preservation and higher strength allowing more repainting. The node adapts its refinement schedule to the detected H3 sampling profile, including base, Turbo, and properly applied PDD configurations. It can also reuse first- and last-frame guidance at the enlarged resolution and optionally divide refinement predictions into overlapping strips to reduce peak video-memory use while retaining the full short-axis context. Native H3 audio is preserved exactly. Its intended place is after the main sampler and before H3 VAE decoding or Star7 Face Repair, providing an integrated resolution-enhancement stage without requiring you to assemble a separate refinement sampler.
MiniMax H3 一键高清放大 - Star7 Input Parameters:
enable_hd
A required BOOLEAN that switches the entire HD stage on or off; the default is True, with options True and False. When True, the node checks the H3 context, evaluates the selected profile and preset, and attempts enlargement and optional refinement. When False, it immediately returns the original audio-video latent unchanged and produces a bypass report, without performing enlargement or refinement. This is useful for comparing the original generation against the enhanced version without disconnecting the node.
sampled_av_latent
A required LATENT connection containing the sampled H3 audio-video result from your main generation stage. This is the material the node enhances: its video component is enlarged and can be refined, while its native audio component is preserved exactly. It must be compatible with the packed H3 audio-video latent format expected by this workflow, rather than an ordinary image or decoded video connection. There is no widget default, numerical range, or fixed list of options. If HD is disabled, or the requested target resolution is no larger than the source, this original latent is returned unchanged.
h3_context
A required STAR7_H3_REFINE_CONTEXT connection supplied by MiniMax H3 All-in-one Conditioning - Star7. It carries information needed to continue processing consistently with your original generation, including the model and positive conditioning; it can also contain first-frame guidance, last-frame guidance, and the Video VAE needed to re-encode those guides at HD resolution. When HD is enabled, missing model or positive conditioning causes an error. If endpoint guides are present and refinement runs, the Video VAE must also be available. There is no widget default, numerical range, or fixed option list; connect the context associated with the video you are enhancing.
upscale_model
A required COMBO selection identifying the learned 3D H3 latent-upscaler checkpoint used for enlargement. Models are expected in models/latent_upscale_models. The default is the pinned FP16 checkpoint designated by the implementation’s _MODEL_NAME constant, and its tooltip states that this default downloads automatically when missing; the supplied excerpt does not reveal its literal filename. Available options are generated dynamically by the model-choice helper, so a complete fixed list cannot be established from the provided code. Use a compatible H3 latent-upscaler checkpoint rather than assuming that an ordinary image-upscaling model can be substituted.
second_pass_lora
A required COMBO selection controlling whether HD refinement uses the first-pass model unchanged or applies a selected model-only LoRA specifically for this second pass. The default is 继承一采, meaning “inherit first pass”; other choices are populated dynamically by the LoRA-choice helper and are not fully listed in the supplied source. Inheritance retains any changes already present on the upstream model rather than resetting it to a clean base model. A separately selected LoRA is applied only when refinement runs, so it has no refinement effect when refinement is disabled or enlargement is skipped. Its effective influence is controlled by second_pass_lora_strength.
second_pass_lora_strength
A required FLOAT setting the model strength of the selected second-pass LoRA. The default is 1.0, the minimum is -100.0, the maximum is 100.0, and the adjustment step is 0.01. This scales the selected LoRA’s contribution during HD refinement; it does not control enlargement size or the fraction of denoising performed. A value of 0.0 also makes sampling-profile detection treat the second-pass LoRA selection as inheritance. Although the interface allows a very broad positive and negative range, that range is not a recommendation for every LoRA: begin with the LoRA’s recommended strength and make small changes to avoid unintended visual changes.
second_pass_attention
A required COMBO selection controlling the attention implementation used for HD refinement. The default is 继承一采, meaning “inherit first pass,” which keeps the existing attention setup. Additional options are provided dynamically by the installation’s attention-backend choice helper; their exact names are not included in the supplied source. Selecting an explicit backend causes the node to install a corresponding model patch for the refinement stage. This is an execution and memory-management choice rather than an artistic strength control, and backend availability can depend on your software and hardware environment.
preset
A required COMBO selection offering 平衡高清 (“Balanced HD,” the default), 高质量 (“High Quality”), 远景小脸 (“Distant Small Faces”), 高速运动 (“Fast Motion”), and 自定义 (“Custom”). Named presets replace target_megapixels, refine_steps, and refine_strength with profile-dependent values, so the values visible in those individual controls are not necessarily the values used during execution. Base and Turbo profiles use separate preset tables, whose exact settings are not included in the supplied excerpt. For recognized PDD presets, the code sets the target to 1.0 megapixel and derives refinement steps and strength from the trained PDD partition. Select 自定义 when you want your individual resolution, step, and strength settings to be used directly, subject to the node’s validation and profile-specific scheduling.
target_megapixels
A required FLOAT specifying the requested output image area in millions of pixels. The default is 1.0, the minimum is 0.20, the maximum is 36.0, and the adjustment step is 0.05. For scale, 1.0 megapixel represents approximately one million pixels per video frame, not a fixed width or height; the report gives the actual resulting dimensions. Higher targets increase the amount of video data processed and can substantially increase memory use and processing time. If the requested target is less than or equal to the source resolution, the node skips both enlargement and refinement and returns the original latent unchanged. Named presets can override this setting, so use 自定义 for direct control.
refine_steps
A required INT controlling the number of refinement sampling steps requested after enlargement. The default is 2, the minimum is 1, the maximum is 50, and the adjustment step is 1. Refinement is intended as a short finishing pass rather than a complete regeneration; additional steps increase sampling work and processing time, but do not guarantee a better result. The implementation clamps submitted values to the 1–50 range, including values from older workflows or API requests. Named presets can replace this setting, and PDD scheduling can adjust the effective step count to match its trained boundaries. To disable refinement, set refine_strength to 0.0 rather than attempting to set the step count to zero.
refine_strength
A required FLOAT specifying the fraction of the denoising path replayed during refinement. The default is 0.25, the minimum is 0.0, the maximum is 0.50, and the adjustment step is 0.01. Lower values keep the enlarged latent closer to its existing appearance, while higher values begin from noisier latents and permit more repainting. A value of 0.0 disables refinement while still allowing learned enlargement if the target exceeds the source resolution. Named presets can override this value, and PDD uses its trained tail boundaries rather than an unrestricted generic schedule. The implementation clamps submitted values to the supported range.
seed
A required INT providing the seed used for refinement noise. The default is 0, the minimum is 0, and the maximum is 18446744073709551615. Keeping it fixed helps you compare settings while holding the requested noise initialization constant; changing it can produce different refinement details when refinement is active. It is not a replacement for the seed used in the original generation and does not independently determine the initial video content. It has no refinement effect when HD is bypassed, enlargement is skipped, or refinement strength is zero; identical seeds also do not guarantee identical results across different software or attention implementations.
enable_tiling
A required BOOLEAN enabling optional strip-based prediction during refinement. The default is False, with options True and False. The node describes this feature as overlapping long-edge prediction strips that reduce peak refinement VRAM use while keeping the full short-axis context. This can help when refining a larger video exceeds available graphics memory, but it should not be assumed to reduce memory use for every stage, such as learned enlargement or final decoding. When enabled, tile_count and tile_overlap configure the strip arrangement.
tile_count
A required INT setting the number of refinement prediction strips when tiling is enabled. The default is 2, the minimum is 2, the maximum is 64, and the adjustment step is 1. More strips divide the long-axis prediction work into smaller portions and can help reduce peak refinement memory demand, although additional divisions can increase processing overhead. This setting is relevant to the optional tiled refinement path, not the output resolution. Start with the default count and increase it only as needed for memory constraints rather than assuming the maximum is preferable.
tile_overlap
A required INT controlling the overlap between neighboring refinement strips when tiling is enabled. The default is 128, the minimum is 32, the maximum is 512, and the adjustment step is 32. Overlap gives adjacent strips shared context instead of treating their boundaries as completely separate regions, helping support continuity across the divided prediction area. Larger overlaps involve more shared processing and can reduce the memory or speed benefit of using smaller strips. The supplied class does not specify the coordinate units for this value, so it should not be interpreted as a confirmed number of decoded output pixels.
MiniMax H3 一键高清放大 - Star7 Output Parameters:
hd_av_latent
A LATENT containing the H3 audio-video result for subsequent processing. When enlargement succeeds, its video component has the increased resolution and includes refinement if that stage was enabled; the native H3 audio is preserved exactly. This remains a compressed internal representation, not a viewable RGB video, so connect it to a compatible H3 VAE decoder or continue through Star7 Face Repair before final decoding. When HD is disabled or the target does not exceed the source resolution, this output is the original latent unchanged. The implementation also tracks existing face-overlay metadata, although the recommended workflow places HD before face repair.
report
A STRING describing what the node actually did. A completed-processing report can include source and output dimensions, actual megapixels, selected upscaler, effective refinement steps and strength, detected sampling profile, HD endpoint-guide status, scheduler and sigma values, second-pass LoRA and attention summaries, tiling information, forwarded face-overlay count, processing times, and confirmation that audio was preserved exactly. Sigma values describe the internal refinement noise schedule and are primarily useful for troubleshooting. If processing is bypassed or skipped, the report explains why, such as HD being disabled or the target being no larger than the source. Check this output when a preset or profile makes the effective settings differ from the values you expected.
MiniMax H3 一键高清放大 - Star7 Usage Tips:
- Place the node after your main H3 sampler and before H3 VAE Decode. If you also use Star7 Face Repair, run HD first and face repair afterward.
- Connect
h3_contextfrom MiniMax H3 All-in-one Conditioning - Star7 associated with the same generation. For first- or last-frame-guided work, ensure that context also includes the Video VAE.
- Use
平衡高清as the default preset starting point. Switch to自定义before tuningtarget_megapixels,refine_steps, orrefine_strengthyourself, because named presets can replace all three settings. - For enlargement with minimal additional repainting, choose
自定义and setrefine_strengthto 0.0. The learned upscaler still runs if the requested resolution exceeds the source. - For a controlled refinement comparison, keep
seedfixed and change one setting at a time. Begin with the default two steps and strength 0.25, then lower strength if the result changes the original appearance more than you want. - Confirm that
target_megapixelsexceeds the source area. For example, a 1280 × 720 source is about 0.92 MP, while a 1920 × 1080 source is about 2.07 MP; a custom 1.0 MP target will not enlarge the latter. - If refinement runs out of graphics memory, enable
enable_tilingand begin withtile_countset to 2 andtile_overlapset to 128. Increase the strip count gradually if necessary; if memory fails during enlargement or decoding instead, reduce the target resolution or address that stage separately. - Leave
second_pass_loraandsecond_pass_attentionon继承一采until the basic workflow works. Then test one refinement-specific change at a time so you can identify its visual and performance effects. - For PDD workflows, apply the PDD model using MiniMax H3 PDD Acc LoRA (Apply), not a normal LoRA loader. The node explicitly states that its HD stage does not require the separate PDD Scheduler node.
- Read
reportafter each test to verify the actual resolution, profile, refinement schedule, and elapsed times rather than relying only on the visible widget values.
MiniMax H3 一键高清放大 - Star7 Common Errors and Solutions:
Star7 H3 HD context is incomplete. Connect H3 context from MiniMax H3 All-in-one Conditioning - Star7.
- Explanation: HD is enabled, but the context does not contain the model or positive conditioning required for processing.
- Solution: Connect the H3 context from MiniMax H3 All-in-one Conditioning - Star7 for the same generation. Check that the upstream conditioning stage completes successfully and that you have not connected an unrelated context.
A PDD file/name was detected upstream, but the trained PDD final-layer head bank was not applied.
- Explanation: The node detects a PDD-related model or filename, but the additional trained components needed for PDD are not properly installed on the model. The full error instructs you to use MiniMax H3 PDD Acc LoRA (Apply) instead of a normal LoRA loader.
- Solution: Replace the ordinary PDD LoRA-loading step with MiniMax H3 PDD Acc LoRA (Apply), then reconnect the resulting model through the appropriate conditioning workflow. You do not need the separate PDD Scheduler node for this HD stage.
Star7 H3 HD received first/last-frame context without the Video VAE needed to re-encode the guides at the HD resolution.
- Explanation: Refinement is attempting to reuse first- or last-frame guidance, but the context lacks the Video VAE required to prepare those guide images for the enlarged video.
- Solution: Supply the appropriate Video VAE through the upstream conditioning workflow and regenerate the context. For an enlargement-only test, select
自定义and setrefine_strengthto 0.0; this avoids the refinement guide-reencoding stage but does not repair the missing context.
Unknown Star7 H3 HD preset: …
- Explanation: The submitted preset does not match one of the supported preset names. This is most likely with manually edited workflows, outdated saved graphs, or API submissions.
- Solution: Choose a current dropdown option:
平衡高清,高质量,远景小脸,高速运动, or自定义. For API requests, send the exact option string rather than an English translation.
H3 latent upscaler produced NaN or Inf
- Explanation: The upscaler produced invalid numerical values rather than a usable enlarged video representation. The node detects this and stops processing.
- Solution: Retry with the pinned default H3 latent-upscaler model and verify that its checkpoint is intact. Confirm that your incoming latent comes from a compatible H3 workflow; if the issue persists, test a smaller target and inspect the ComfyUI console for related model or device errors.
Upscaler checkpoint missing or automatic download failed
- Explanation: This is a possible setup failure rather than an exact error string shown in the supplied source. The selected checkpoint may be unavailable in
models/latent_upscale_models, or the default checkpoint’s automatic download may fail. - Solution: Check your network connection, available disk space, and folder permissions. Use the extension’s supported installation instructions to obtain the intended checkpoint, and select a compatible H3 latent-upscaler model after refreshing or restarting ComfyUI.
CUDA out of memory or another device-memory allocation error
- Explanation: This is a possible runtime error whose exact wording depends on your environment. Enlargement, refinement, or decoding can exceed available graphics memory, especially at high megapixel targets.
- Solution: Select
自定义and lowertarget_megapixels. If the failure occurs during refinement, enable tiling and gradually increasetile_count. Close other GPU-heavy applications, and remember that refinement tiling does not guarantee lower memory use during enlargement or decoding.
Attention backend unavailable or failed to initialize
- Explanation: This is a possible environment-dependent failure, not a verified literal error message from the supplied excerpt. An explicitly selected second-pass attention backend may require software or hardware support that is not available.
- Solution: Return
second_pass_attentionto继承一采and retry. If you need a specific backend, check its installation requirements and confirm that your ComfyUI environment supports it.
Star7 H3 HD skipped | … | target <= source; first-pass latent returned unchanged
- Explanation: This report is not an error. The requested target does not exceed the source resolution, so the node intentionally performs neither enlargement nor refinement.
- Solution: Select
自定义and settarget_megapixelsabove the source image area. Check whether a named preset had replaced your intended target.
Star7 H3 HD bypassed | HD second pass disabled | latent unchanged
- Explanation: This report is not an error.
enable_hdis False, so the node returns the original latent without processing. - Solution: Set
enable_hdto True to enable enhancement. If you only want to disable refinement while keeping enlargement, leave HD enabled, select自定义, and setrefine_strengthto 0.0.
