13:12
Flux
The video argues that the ultra-realistic Flux AI images and talking-head videos flooding Reddit and X in mid-2024 aren't just the stock Flux model at work — the realism comes from a specific 'flux realism' LoRA and a lowered guidance scale, and the presenter demonstrates that this look can be reproduced end-to-end via fal.ai for image generation and Runway Gen-3 Alpha for animation, while suspecting the most impressive viral clips were cherry-picked from multiple rerolls.
AI images generated with Flux have become realistic enough that scrolling past them on Instagram, viewers often wouldn't know they're AI-generated.
The imperfect, off-center 'amateur snapshot' composition of these images (rather than polished professional-photo framing) is part of what makes them read as real, per the video.
Full-body shots are the weak point where proportions can look 'wonky,' though the video says a few rerolls usually fixes it.
People on X began animating these realistic Flux images into talking-head videos (e.g. with Luma's Dream Machine), amplifying how convincing they looked.
The presenter's own Flux generations (run through Glif) looked noticeably less realistic than the viral Reddit examples — the skin had a 'plastic shininess.'
The realism gap is traced to a LoRA (Low-Rank Adapter): a small fine-tuning add-on layered on the Flux foundation model that specifically improves skin, hair, and wrinkle rendering.
Glif doesn't support adding LoRAs at all (confirmed under Advanced controls and Add Block), which the video says is why the presenter's images lacked the extra realism.
Two viable ways to use LoRAs with Flux, per the video: ComfyUI (powerful but complex, described as 'spaghetti bowl' node graphs) or fal.ai (a cloud inference service, described as easier).
fal.ai hosts flux-realism, flux-1-pro, and flux-1-dev; the video reports flux-realism generations cost about $0.32 each, flux-pro about $0.05, and new signups got $2 in free credit at time of recording.
Lowering the guidance/CFG scale from fal.ai's default 3.5 to about 2 (keeping inference steps at 28) was the presenter's found 'sweet spot' for realistic, non-shiny skin.
The presenter's recommended pipeline: generate via fal.ai's flux-realism LoRA with guidance scale ~2, then animate the image in Runway Gen-3 Alpha, which the video says produced better results than Luma Dream Machine (whose output had the face go 'wonky' toward the end).
The presenter believes the most impressive viral 'AI person talking' videos on X were probably cherry-picked from several reroll attempts rather than one-shot outputs, since his own single attempts showed visible flaws (a floating microphone, a microphone held unnaturally still despite head movement, wonky fingers).
Flux — The foundational text-to-image AI model that generates the base image before any LoRA fine-tuning is applied, per the video. Apply: Run prompts through Flux (via Glif's free Pro access or fal.ai) as the starting point before layering on a realism LoRA.
LoRA (Low-Rank Adapter) — Per the Perplexity explanation quoted in the video, a small (2–500MB) add-on that fine-tunes a base model on specific concepts, styles, or characters without retraining the whole model. Apply: Layer a LoRA on top of Flux (via ComfyUI or fal.ai) to target improvements like style, character consistency, or image realism/quality.
Flux Realism LoRA (from Excel lab/XLabs) — A specific LoRA that, per the video, adjusts skin, hair, and wrinkle rendering to make Flux outputs look more photorealistic. Apply: Select the flux-realism model on fal.ai and pair it with a lowered guidance scale (~2) to reproduce the ultra-realistic look seen in the viral Reddit/X images.
Glif (glif.app) — A workflow-builder app the presenter used to run Flux, including free access to the Flux Pro version, but which the video shows has no option to add LoRAs. Apply: Use Glif for quick, free Flux generations when LoRA-level realism isn't required; per the video it can't be used when a realism LoRA is needed since it's unsupported there.
ComfyUI — A node-based ('spaghetti bowl') AI workflow tool that can run LoRAs, described in the video as powerful but complex and easy to get lost in past three or four blocks. Apply: Build a ComfyUI graph including the Flux model and a LoRA node for more dialed-in control, per the video, if willing to handle the added complexity.
fal.ai — A cloud inference service, compared in the video to Replicate or Hugging Face Spaces, that lets users run Flux and Flux LoRAs like flux-realism on fal's own compute. Apply: Sign up (new accounts reportedly got $2 in credit per the video), go to fal.ai/models, pick flux-realism, and run generations at roughly $0.32 each (vs about $0.05 for flux-pro).
Guidance scale / CFG scale tuning — An adjustable generation parameter that the video says defaults to 3.5 on fal.ai but produces shiny, plasticky, unrealistic skin at that setting. Apply: Lower the guidance/CFG scale to about 2 (keeping inference steps at 28) to get the most realistic-looking output, per the presenter's own testing.
Runway ML Gen-3 Alpha — An image-to-video tool the presenter used to animate the realistic Flux image into a talking-head video clip. Apply: Upload the generated image, crop to frame the subject fully, paste the (shortened, under-500-character) original prompt, and generate; per the video it produced more realistic results than Luma's Dream Machine, though with artifacts like a floating microphone or slightly wonky fingers.
Luma Dream Machine — An alternative image-to-video generator the presenter tested with the same image and prompt as Runway. Apply: Feed in the same realistic image and prompt as an alternative to Runway; per the video the results were noticeably worse, with the face going 'wonky' toward the end of the clip.
The realism difference between 'good' and 'bad' Flux outputs isn't presented as random luck — the video isolates it to two concrete, controllable levers: which LoRA is loaded and the guidance-scale value, both tested directly by the presenter.
Counterintuitively, off-center/imperfect composition — the opposite of polished professional photography — is framed as the actual realism cue, since a too-perfect AI image now reads as more suspicious than an 'amateur phone snapshot' framing.
Tool choice, not skill or prompting, was gating the presenter's results: Glif structurally cannot add LoRAs, so no amount of prompt refinement inside Glif could reach LoRA-level realism — a limitation the video verifies by checking Glif's UI directly rather than assuming it.
The presenter reverse-engineers a 'cherry-picked' production hypothesis for viral AI-video clips purely from his own reroll experience — inferring an unseen production process (many attempts, best one shared) from the gap between his single-take output quality and what circulates online.
«man these AI generated images have been really good lately»
— 00:00
«here's another one I mean as you can see it's getting harder and harder to tell when an image was generated with AI»
— 00:49
«I think the fact that they're not like perfectly composed like they don't look like a professional photographer took them is sort of what gives them that feeling of all right this looks like just a random snapshot that somebody took it looks real»
— 01:01
«flux is the foundational image model which generates the image the Laura is like some extra sort of fine-tuning information on top of that training»
— 03:56
«allora is used to train the model on specific Concepts Styles or characters allowing for targeted improvements in image quality style specificity or character consistency»
— 04:10
«you can leave the number of inference steps at 28 here but if you leave the guidance scale at the default 3.5 it actually doesn't look that great it starts to look shiny and plasticky and more unrealistic I found the sweet spot to be about two»
— 08:24
«I'm fairly certain that most of the videos that you're seeing on X where somebody took a ultra realistic AI generated human and made a video of that person speaking it was probably a little cherry-picked they probably had to do a few rerolls»
— 11:19
Reception
Generally positive reception from Flux enthusiasts and beginners, though critics found the video too long and noted legitimate concerns about AI limitations, platform issues, and preferences for local solutions.
A practically useful teardown that isolates the specific technical levers (LoRA choice, guidance-scale tuning) behind Flux's viral-looking realism, honestly showing that the presenter's own reproductions don't match the polish of the examples he opens with.
13:12