A user is asking about the best models, quantization, and workflow solutions for local Text-to-Video and Image-to-Video generation with ComfyUI on an RTX 5070 Ti (16GB GDDR7 VRAM) + Ryzen 5 7600X + 32GB DDR5 setup.
The author tested five local MiniMax H3 recipes using the same 15-second three-person alien village scene, with generation times ranging from 95 seconds to 843 seconds.
The author ran MiniMax H3 Singularity V2 Pruned INT8 ConvRot R2 with ComfyUI on an RTX 3060 12GB, generating an approximately 10-second video at 1536×864 resolution, 243 frames, and 4-step sampling, taking 608 seconds.
On the evening of October 25, the author used Meta's AI agent Muse to generate a 30-second vertical video for the self-made "Tianxia Buxian" series. The first three requests produced one 10-second trial and two 30-second versions; the last one was checked and found to be 720×1280, 24fps, 30.000 seconds, with an audio track.
The author tried out the ControlNet-LLLite for Anima in ComfyUI, which can use 3D poses to specify character actions and region control to specify the generation range.
A user is trying to generate loop videos in Minimax using the same first-and-last-frame method, but the motion always "starts" at the 1st second and decelerates in the last 1 second, making it impossible to achieve a seamless constant-speed transition. Taking a 15-second hammering loop as an example, the actual result is 1 second of startup, 13 seconds of steady motion, and 1 second of wind-down before jumping back to the beginning. The user has tried prompts such as "video is a clip of the action segment" with no success, and is now seeking a working workflow, LoRA, or prompt solution.
A father plans to use ComfyUI on Runpod to make a short film in which his daughter, who loves Game of Thrones, marries her partner in a medieval castle surrounded by family. He says he is new to image generation and workflows, is learning by trial and error, and expects to need DaVinci for editing later, and he is publicly asking for advice.
AIwood Lab found during H3 VBVR testing that loading VBVR did not meet expectations and was even worse; after switching to the older ComfyUI 0.38 environment and comparing with 0.42, noticeable differences appeared in scenes such as the chef.
On a machine with an RTX 3060 Ti (8GB VRAM) and 16GB of RAM, the author used Qwen-Image-2.1 to generate an image of an underwater train carriage, then used MiniMax H3 together with Veda and Endless to extend it segment by segment into a 25-second video.
Before the PR video is finished, watch the exported file in 3 passes: first watch the content from a first-time viewer's perspective, then look at the text and the lower part of the frame, and finally listen to the audio only. Classify the problems you find as "unclear communication," "product distortion," or "unreadable/inaudible," and note the seconds and change points in a revision note. Change only one place at a time; after each change, bump the version number, re-export, and keep the old version. When done, archive the video together with the work description note, labeled "Fictional case · self-produced."
The author used Animagine XL 4.0 Opt in ComfyUI to generate a Halloween little devil with silver hair, red eyes, short black horns, wearing a black-and-orange outfit and looking up at the camera; the resulting image is 1280×1024.
Creator Naoya disclosed the workflow for making the 18-minute 24-second, 4K, 129-shot Sekigahara-themed historical film "Shimazu's Retreat" using tools such as MiniMax H3. The warlords' lines in the film were switched to Japanese voice acting; the author wrote the lines in hiragana without kanji, wrapped them in dedicated tags, and added an instruction to close the mouth after speaking, to reduce misreading and text appearing in the frame. After watching all 129 shots one by one, they were reshot, fixing issues such as helmets disappearing, two people speaking at the same time, reversed forms of address, and backgrounds that looked like a modern park; the ending was changed to a 34-second fade-out of the theme song.
The author built a MiniMax H3 workflow in ComfyUI, combining SelfLift progressive resolution sampling with a DLSS 5 upscaling node to output 2K or 4K video on an RTX 3060 with 6GB of VRAM. In their 6GB VRAM environment, generating and upscaling to 4K took about 35 minutes, with parameters of 8 steps and 0.6 megapixels, all without additional model training; both the workflow and the DLSS5 DLL download links have been released.
Why it matters: The author tested the full pipeline from low-resolution progressive sampling to 4K upscaling on a 6GB VRAM graphics card and provided the time taken and parameters, as a reference for users with low-spec devices.
Pollo AI's Recast feature can reference the original video's motion and camera movement while replacing the characters and background; the author used it to convert a live-action two-person dance into original Japanese-style animation characters. The same material was generated at 480p and 720p respectively; 480p consumed 235 credits and 720p consumed 505 credits, a difference of 270 credits. The 720p details are slightly clearer, but the 480p result already exceeded expectations, and the author recommends trying 480p first.
The core difference between Higgsfield and Runway is: the former is a model aggregator, with a single subscription including about twenty video models such as Seedance, Kling, Veo, and Wan plus Cinema Studio 4.0, suited to generating shots from scratch; the latter uses its self-developed Gen-4.5 for generation and Aleph 2.0 for refining existing videos, with the Max tier supporting ProRes and HDR export.
Claude set up a ComfyUI workflow on an M1/16GB Mac; after adjusting launch options, single-step generation went from 10 minutes to about 10 seconds, and they combined processes including a watercolor LoRA, black-and-white manga switching, inpainting, and upscaling.
The author used Animagine XL 4.0 Opt in ComfyUI to generate an example of a female rock singer holding a mic, leaning forward, and belting, at a resolution of 1248×832, Seed 2026101857, Steps 28, CFG 5, euler_ancestral. The accompanying JSON uses only 7 standard nodes and requires no input image, LoRA, ControlNet, or custom nodes; two comparison JSONs are also provided, one removing the forward lean and one removing the singing expression.
A user tested a high token limit on MOSS SFT with a prompt of about 1.3k tokens, combined with a 4-steps turbo LoRA, 4MP resolution, and Euler/simple sampling, to generate a photo containing 12 different people. Each character's appearance, clothing, and props remained consistent, with no color bleeding or duplication.
In a local MiniMax H3 image-to-video workflow, MOSS-Image (5B) native 4MP output combined with a simple 3–4 word prompt generated a roughly 3-second clip, with detail retention and temporal consistency exceeding expectations. The suede jacket texture, skin pores, chains, tweed and thin frames all showed no texture sliding, and the second video also preserved the stylized elongated neck proportions rather than forcibly correcting them. The test was based on an NVIDIA RTX PRO 6000, with the built-in MOSS template in ComfyUI as the workflow.
The author shares hands-on results from training a character LoRA on MOSS Image, saying that using Automagic3 with Sigmoid and the High setting, a learning rate starting at 0.0005 with a minimum learning rate set, and not excluding block 22 produced good results every time.
Aizawa Amu posted 5 illustrations on X from the morning of October 24 to late night on October 25, and freely distributed the corresponding ChatGPT image-generation prompts, covering five scenes: morning, daytime, evening, night, and late night. The prompts require attaching a reference image of an adult character aged 21 or over that the user has the right to use, preserving facial features, hairstyle, hair color, and eye color, and outputting a 1536:1024 landscape single-scene anime-style illustration. The author also suggests changing only one detail, for example swapping the daytime plaid fabric sample for a polka-dot pattern.
A user trained a style LoRA for Qwen-image with ai-toolkit on a Colab 80GB A100, but the output backgrounds were too simple, random people and duplicate characters appeared, and the style drifted. The configuration was rank 16, 1024px, bf16, adamw8bit; a learning rate of 1e-4 for 500 steps had no effect, so it was changed to 3e-4 and trained to 2k steps, and a 12-image dataset still showed no improvement. The user plans to expand to 30–40 images.
Using Animagine XL 4.0 Opt with Control-LoRA Canny rank128 contour guidance, a 1248×832 landscape composition of an elf drawing a bow with Seed 2026101813 was generated in ComfyUI, and the contour PNG and a JSON containing the node configuration were released.
A collection of 100 AI beauty image-generation prompts for tools such as ChatGPT and Gemini has been released, all in directly copyable JSON format. Each prompt set includes prompt, negative_prompt, and aspect_ratio fields, and provides 1:1, 9:16, and 1280×670 sizes plus background and mood adaptation versions. The themes focus mainly on beauties in spectacular and everyday scenes around the world, such as Argentine salt flats and Kyoto riverside terraces.
The author used ChatGPT to generate Halloween-themed images for original characters Popo, Bobo, and Shuka, but the character designs frequently went wrong: Bobo was given hands and feet, Shuka was put in a crooked hat, and inexplicable elements such as "Maple Valley Station" appeared. The author also shows successful cases using watercolor-style prompts that blend watercolor bleeding effects into small objects and the ground, making the white-dress setting for non-human characters work. The author also suggests that when fixing failed images, regenerating from the original reference image has a higher success rate than repeated local edits.
A user says they installed ComfyUI on their PC the previous night and downloaded the dependencies needed to generate video, then received a low disk space warning, but on checking found the app occupies only 500MB. Confused about where ComfyUI's downloaded files are actually stored, they posted for help.
A free comparison page summarizing 18 ComfyUI configurations has launched, allowing filtering by GPU, model and precision and calculating seconds saved. On the 3090 Ti, MOSS BF16 drops from 71.98 seconds to 45.40 seconds, while MOSS INT8 only goes from 27.31 seconds to 27.23 seconds; H3 INT8 ConvRot drops from 155.07 seconds to 118.21 seconds.
A 3-step method for AI music beginners: first decide on a theme for the song, then connect the protagonist, setting, and emotion into a beginning–middle–end narrative, and finally hand the content to ChatGPT to generate production notes. The production notes include a 3-line story, 5 original song titles, key points for verses and chorus, a mood description of 80 characters or fewer, and a self-check list before release; ChatGPT only organizes the narrative and production notes, and a separate music-generation tool is still needed to complete the audio, and the terms of service must be checked before public release or commercial use.
A 49-year-old creator switched their AI video generation model from Adobe Firefly Video to Google Veo 3.1 Fast and completed a kitten short film matching the specified first and last frames without any redo.
A user shared a Halloween-themed ComfyUI creation workflow that uses MOSS-Image to generate portraits with extremely ugly faces, then uses MiniMax H3 Ref2VA's reference generation capability to transfer the same face onto a full-body model. The example prompt describes features such as warts, pimples, buck teeth, and a double chin in detail, and provides a continuous shot design from a full-body push-in to a facial close-up.
Higgsfield released a complete guide to AI filmmaking, using a 10-second seaside station scene as an example to demonstrate generating characters with Seedream 5.0 Pro, generating scene reference images with Soul Cinema, then having Supercomputer expand the storyboard prompts and Cinema Studio 4.0 generate a multi-shot clip with sound in one pass in multishot mode.
The author used MOSS-Image's JSON bbox specification in ComfyUI to arrange each person individually, testing three sets of scenes with 3, 6, and 10 people; with 10 people, the number of people, arrangement order, and each person's hair color and clothing basically matched the specification, with only one person missing eyes when the LoRA was disabled.
The author has released a set of monolid (single eyelid) prompts for AI image generation, requiring both eyes to be natural monolids with no upper eyelid crease, along with correction prompts for when double eyelids are generated. In testing, ChatGPT's regular Chat vaguely showed monolids, while ChatGPT Work 6Astra instead generated double eyelids; the author says the higher-tier model feels like it has a higher success rate but cannot guarantee it.
The user runs flux2Klein_9bKv_pruned_fp8 and flux-2-klein-9b-fp8 (with qwen_3_8b_fp8mixed and flux2-vae) on a 3090ti to generate consistent characters. A single reference image without LoRA takes about 30 seconds, and two reference images take about 2 minutes, but once a LoRA is loaded it hangs at KSampler 70%, with GPU usage as high as 96-97%.
A ComfyUI user gets only 1 to 4 fps when manually masking in the Load Image node, with a configuration of RTX 4090 plus Ryzen 9900X, and believes it is not a hardware bottleneck. The user says it was smooth before a major update earlier this year, currently uses ComfyUI_frontend v1.53.6 and PyTorch 2.9.1+cu130, has enabled Firefox hardware acceleration, and also uses the SAM text selection method.
MOSS-Image was tested at native 4MP with Krea 2 prompts, taking about 25 seconds per image and using 16-24GB VRAM on an RTX PRO 6000, based on the official built-in ComfyUI template workflow, with all results being unselected first runs. The model responds well to casual flash-photography texture, subtle skin texture, and stylized proportions, but seed variance is very large—each seed significantly changes the composition, so seeds must be filtered when a specific composition is needed.
A user making posters and infographics in ComfyUI wants to use the native Bounding Box capabilities of multiple models for layout and local inpainting, but is paralyzed by analysis due to too many choices, and asks the community for experience and tips from those who have actually used multiple models or nodes.
The key to creating a rainy-day atmosphere with AI illustration is to manage the direction of light through perspective and to add particle impurities such as raindrops as moderate noise. Using car headlights backlighting a rain curtain as an example, the author points out that the clearer the light band and directionality the better, but the actual effect is only light faintly penetrating from behind. They also emphasize that expression and emotion should come before the weather setting—a smiling face with rain looks eerie.
A user asks the ComfyUI community for help with comic-style redrawing, encountering two core problems: comics contain text, and direct redrawing deforms the text; a single image may contain multiple characters, making it hard to keep their expressions basically consistent with the original. They have tried some methods but can at best change colors and cannot alter the original line art, so they seek workflow or model suggestions.
A user's LoRA training for MOSS in AI toolkit completely collapsed, prompting them to question whether new models (Qwen 2.1, LLaMA 2, MOSS) should use the shift or sigmoid parameter. The user mentions that sigmoid works on Krea 2 and Qwen 2.1, and asks whether to start with shift balanced and switch to shift low noise later in training.